An end-to-end generative scheduling system and method for Ascend container cloud environment

Abstract

The application discloses an end-to-end generative scheduling system and method for Ascend container cloud environment, a scheduling scene analysis module generates a time sequence index heat map and extracts a visual semantic feature sequence; a scheduling strategy planning module receives user intention text and the visual semantic feature sequence, queries a meta directory of a scheduling interface engine module, retrieves similar historical cases from a scheduling experience knowledge sample library, constructs a structured task prompt word, and guides a generative artificial intelligence to output executable interface-level scheduling strategy code; a scheduling strategy evaluation module performs rule evaluation and runtime evaluation on the interface-level scheduling strategy code; the scheduling interface engine module provides a standardized task scheduling and Ascend container cloud environment resource management interface library and meta directory; and a data enhancement module generates an expanded sample set and an image-text sample pair to train the generative artificial intelligence in the scheduling strategy planning module. The application improves the flexibility and stability of scheduling decision of the Ascend container cloud environment.

Description

Technical Field

[0001] This invention relates to the field of container cloud task scheduling and artificial intelligence, specifically to an end-to-end generative scheduling system and method for the Ascend container cloud environment. Background Technology

[0002] Container cloud, with its lightweight, highly elastic, and agile deployment characteristics, has become the mainstream infrastructure supporting the training, inference, and collaborative services of geospatial application models and large-scale geographic information models. Container task scheduling, as a crucial support for cloud computing resource management, aims to dynamically match task instances with containers based on user needs, task characteristics, and the real-time status of cluster resources. Current mainstream methods rely on static configuration and manual templates, which are not only costly to maintain but also difficult to adapt to the dynamic needs of service collaboration, cross-access, and real-time interaction between geographic models. Furthermore, most systems lack effective scheduling metric evaluation and feedback mechanisms, making them prone to scheduling imbalances in large-scale heterogeneous clusters where geospatial application models and large-scale geographic information models are deployed together. To achieve complete autonomy and control over computing resources, building a container cloud environment based on the Ascend platform has become the industry's preferred choice. However, unlike the scheduling of general-purpose computing resources such as CPUs and GPUs, container task scheduling for the Ascend container cloud environment faces greater challenges. Because the Ascend series NPUs are based on the Da Vinci architecture, their computing units and hardware accelerators differ from traditional processors. On the one hand, scheduling strategies optimized for general platforms are difficult to apply directly; on the other hand, they place higher demands on fine-grained management of heterogeneous resources and the effective achievement of diverse service level goals. In recent years, artificial intelligence technology, with its autonomous decision-making and knowledge learning capabilities, has provided a new path to overcome the bottlenecks of traditional scheduling technologies. In particular, the multimodal transformation and knowledge generation capabilities of generative artificial intelligence make it possible to optimize dynamic scheduling strategies in the Ascend container cloud environment. However, current applications still face significant challenges: first, the lack of a domain knowledge system deeply integrated with the Ascend container cloud environment leads to deviations between generated strategies and actual needs; second, the generated results are mostly decision suggestions, requiring manual secondary development into executable code; and third, the lack of effective verification and feedback mechanisms results in risks such as logical conflicts and violations of requirements. In summary, the container cloud scheduling field, especially for the domestically produced Ascend container cloud computing environment, still lacks a highly reliable end-to-end generative scheduling solution covering "intent understanding - strategy generation - result verification - feedback optimization." Summary of the Invention

[0003] The purpose of this invention is to provide an end-to-end generative scheduling system and method for the Ascend container cloud environment, so as to enhance the understanding of scheduling intentions and the dynamic adaptability to the Ascend container cloud environment, and improve the performance of container task scheduling in the domestic environment.

[0004] The technical solution for implementing this invention is: an end-to-end generative scheduling system for the Ascend container cloud environment, comprising a scheduling scenario analysis module, a scheduling strategy planning module, a scheduling strategy evaluation module, a scheduling interface engine module, and a data enhancement module connected in sequence, wherein:

[0005] The scheduling scenario analysis module is used to extract resource indicators and task indicators with time-series correlation from the Ascend container cloud environment, generate a time-series indicator heatmap composed of resource indicator heatmap sequences and task indicator heatmap sequences, and extract visual semantic feature sequences from the time-series indicator heatmap and output them to the scheduling strategy planning module.

[0006] The scheduling strategy planning module is used to receive user intent text and visual semantic feature sequences, query the meta directory of the scheduling interface engine module, retrieve similar historical cases from the scheduling experience knowledge sample library by fusing task semantics and resource matching similarity function, and construct structured task prompt words by combining domain knowledge obtained from multi-level knowledge support system, so as to guide the generative artificial intelligence to output executable interface-level scheduling strategy code, and output it to the scheduling strategy evaluation module.

[0007] The scheduling strategy evaluation module is used to perform rule evaluation and runtime evaluation on the received interface-level scheduling strategy code, and send the error feedback or runtime feedback information corresponding to the strategy that fails the evaluation back to the scheduling strategy planning module to regenerate the strategy. At the same time, the scheduling strategy code that passes the evaluation is archived to the scheduling experience knowledge sample library.

[0008] The scheduling interface engine module is used to provide a standardized task scheduling and Ascend container cloud environment resource management interface library and meta directory for the scheduling strategy planning module. The interface library includes a scheduling environment awareness interface and a scheduling action reasoning interface. The meta directory includes interface name, interface type, function description and method list.

[0009] The data augmentation module is used to process the original data through text data augmentation strategies and image data augmentation strategies to generate an expanded sample set and image-text sample pairs to train the generative artificial intelligence in the scheduling strategy planning module.

[0010] Furthermore, the scheduling scenario analysis module includes a time-series indicator heatmap generation submodule and a visual semantic feature sequence extraction submodule, wherein:

[0011] The time-series indicator heatmap generation submodule is used to generate a time-series indicator heatmap composed of a resource indicator heatmap sequence and a task indicator heatmap sequence. The specific method is as follows:

[0012] Based on the resource remaining rate of computing resource nodes The first linear mapping formula is used to convert the values ​​into grayscale values, generating a series of resource indicator heatmaps.

[0013]

[0014] Task occupancy rate based on computing resource nodes The values ​​are converted into orange series values ​​through the second linear mapping formula, generating a heatmap sequence of task indicators.

[0015]

[0016] in, The function rounds a floating-point number to an integer value. Indicates the computing resource node number. Indicates the computing resource type number, Indicates the containerized task type label;

[0017] The visual semantic feature sequence extraction submodule is used to extract visual semantic feature sequences, and the specific method is as follows:

[0018] A pre-trained visual encoder is used to extract high-order visual features from a time-series index heatmap;

[0019] By using a multilayer perceptron consisting of two fully connected networks connected in sequence, each configured with a GeLU activation layer and a LayerNorm normalization layer, linear projection is performed to align high-order visual features to the text embedding space of generative artificial intelligence, forming a sequence of visual semantic features.

[0020] Furthermore, in the scheduling interface engine module,

[0021] Both the scheduling environment awareness interface and the scheduling action reasoning interface adopt a three-tier inheritance architecture including an abstract base class, an abstract class, and an implementation class.

[0022] The abstract classes of the scheduling environment awareness interface include a perception class based on vector representation, a perception class based on image representation, and a perception class based on graph representation.

[0023] The abstract classes of the scheduling action reasoning interface include heuristic reasoning classes, swarm intelligence-based reasoning classes, deep learning-based reasoning classes, and deep reinforcement learning-based reasoning classes.

[0024] Furthermore, the scheduling strategy planning module constructs task prompt words, including:

[0025] The API definition section uses semantic analysis to compare the user intent text with the functional description text of the interface in the meta directory of the scheduling interface engine module, and filters out interface information with a matching degree higher than a preset threshold.

[0026] The system prompts are based on a multi-layered knowledge support system, including Ascend Container Cloud architecture expertise, infrastructure configuration information, Ascend Container Cloud environment monitoring data, and task dependencies.

[0027] The context example section calculates the semantic similarity of tasks and the similarity of resource requirements, and retrieves the most similar historical scheduling cases from the scheduling experience knowledge sample library.

[0028]

[0029] in, For the current task, For the l-th historical scheduling task, for User intent text embedding vector, for User intent text embedding vector, The formula for calculating resource similarity based on resource demand vectors is as follows:

[0030]

[0031] in Temperature is a parameter used to describe the importance of resource similarity. For the Sigmoid function, and These are the standardized resource requirement vectors for the current task and the historically scheduled tasks, respectively.

[0032] The task instruction section includes a document string summarized from the user intent text and a thought chain instruction that guides distributed reasoning. The document string summarizes the core objectives, constraints, and key execution requirements of the current scheduled task from the user intent text.

[0033] The feedback correction instruction section incorporates the feedback information and the original scheduling strategy code into the task prompt when it receives feedback information from the scheduling strategy evaluation module.

[0034] Furthermore, the scheduling strategy evaluation module performs rule evaluation and runtime evaluation, wherein:

[0035] The rule evaluation includes interface validation rule evaluation, affinity validation rule evaluation, and conflict validation rule evaluation, wherein:

[0036] Interface validation rule evaluation is used to verify whether the output data type of the scheduling environment awareness interface matches the input data type of the scheduling action reasoning interface in the interface-level scheduling strategy code, and whether the semantic constraints of the interface combination are met.

[0037] Affinity verification rule evaluation is used to verify whether the scheduling mapping relationship defined by the interface-level scheduling strategy code conforms to the node affinity constraint and the task affinity constraint;

[0038] The conflict verification rule evaluation is used to verify whether the scheduling mapping relationship defined by the interface-level scheduling policy code avoids resource contention or exclusive conflicts.

[0039] The runtime evaluation includes task-level error tolerance evaluation and cluster-level error tolerance evaluation, wherein:

[0040] Task-level error tolerance assessment uses standardized response time as the indicator for latency-sensitive tasks and standardized scheduling overhead as the indicator for cost-driven tasks.

[0041] Cluster-level error tolerance assessment uses the standard deviation of resource utilization distribution among containers as an indicator.

[0042] Furthermore, in the data enhancement module,

[0043] The text data augmentation strategies include synonym replacement, noise injection, and semantic invariant rewriting of user intent text.

[0044] The image data enhancement strategy includes random cropping, noise injection, and random occlusion of the time-series indicator heatmap.

[0045] The image-text sample pairs include positive sample pairs and negative sample pairs; the positive sample pairs consist of the original time-series indicator heatmap and its corresponding enhanced image, user intent text and its enhanced text, and scheduling strategy code; the negative sample pairs consist of unrelated time-series indicator heatmaps and text data.

[0046] Furthermore, the system is deployed on an Ascend container cloud cluster that includes Ascend NPU and Kunpeng CPU, and the container applications running in the Ascend container cloud environment include inference and training tasks for geospatial application models and large geographic information models.

[0047] Based on the differences in resource preference types, the geospatial application model is divided into CPU-intensive, memory-intensive, disk-intensive, bandwidth-intensive, and resource-inefficient types.

[0048] An end-to-end generative scheduling method for the Ascend container cloud environment, used in the end-to-end generative scheduling system for the Ascend container cloud environment, the method includes:

[0049] By using the scheduling scenario analysis module, time-series correlation indicators are extracted from the Ascend container cloud environment, a time-series indicator heatmap is generated, and a visual semantic feature sequence is extracted.

[0050] The scheduling strategy planning module receives user intent text and visual semantic feature sequences, queries the metadirectory of the scheduling interface engine module, and guides generative artificial intelligence to generate interface-level scheduling strategy code.

[0051] The scheduling strategy evaluation module performs rule evaluation and runtime evaluation on the interface-level scheduling strategy code. If the evaluation fails, feedback information is generated and sent back to the scheduling strategy planning module to regenerate the strategy. If the evaluation passes, the scheduling strategy code is archived and deployed for execution.

[0052] Through the scheduling interface engine module, a standardized task scheduling and Ascend Container Cloud environment cluster resource management interface library and meta-directory are provided for the two-stage scheduling planning of environment perception and action reasoning in the scheduling strategy planning module, supporting dynamic plug-and-play integration of existing scheduling and newly developed scheduling interfaces.

[0053] The data augmentation module enhances the original text and image data, generating an expanded sample set and image-text sample pairs for training generative artificial intelligence.

[0054] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the end-to-end generative scheduling method for the Ascend container cloud environment.

[0055] A computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the end-to-end generative scheduling method for the Ascend container cloud environment.

[0056] Compared with existing technologies, this invention has the following significant advantages: 1) It constructs a generative scheduling system for the Ascend container cloud environment. Relying on interface-level code directly generated by generative artificial intelligence, it achieves an end-to-end closed loop from scheduling scenario analysis to policy execution, effectively improving the flexibility and stability of scheduling decisions in the Ascend container cloud environment. 2) The end-to-end generative scheduling system integrates an interface engine-driven container scheduling retrieval enhancement link and a policy evaluation feedback mechanism. By dynamically reflecting on and summarizing scheduling results, it significantly improves the adaptability to highly random environmental changes. 3) Based on the above effects, this invention demonstrates superior scheduling performance for tasks such as geospatial application models and large-scale geographic information models on domestically produced computing power represented by the Ascend chip. This invention has a standardized operating procedure and strong scalability, possessing outstanding practical value and promising prospects for widespread application. Attached Figure Description

[0057] Figure 1 This is an architecture diagram of the end-to-end generative scheduling system for the Ascend container cloud environment according to the present invention.

[0058] Figure 2 This is a flowchart of the end-to-end generative scheduling method for the Ascend container cloud environment according to the present invention.

[0059] Figure 3 This is a schematic diagram of the scheduling scenario analysis module.

[0060] Figure 4 This is the architecture diagram of the scheduling interface engine module.

[0061] Figure 5 This is a structural diagram of the task prompt words in the scheduling strategy planning module.

[0062] Figure 6 This is a schematic diagram of the positive and negative image-text pairs in the data augmentation module. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] like Figure 1 and 2 As shown, this invention discloses an end-to-end generative scheduling system for the Ascend container cloud environment, comprising a scheduling scenario analysis module, a scheduling interface engine module, a scheduling strategy planning module, a scheduling strategy evaluation module, a data augmentation module, and an Ascend container cloud cluster. These modules work collaboratively to achieve end-to-end closed-loop scheduling integrating scheduling representation reasoning, scheduling execution observation, and scheduling feedback compensation.

[0065] (I) Scheduling Scenario Analysis Module

[0066] The scheduling scenario analysis module extracts time-series correlated indicator data from the Ascend Container Cloud environment as input. Through the aggregation and visualization of these time-series indicators, it outputs a heatmap of time-series indicators showing the current resource usage status and task distribution in the Ascend Container Cloud environment to the generative scheduling system. This heatmap is then transformed into a visual semantic feature sequence to enhance the model's understanding of the scheduling environment. Specifically, the time-series indicator heatmap consists of a set of time-continuous resource indicator heatmaps and a set of task indicator heatmaps. The extracted visual semantic feature sequence is used to form a multimodal input together with text prompts, guiding generative artificial intelligence in scheduling strategy planning.

[0067] The scheduling scenario analysis module includes two sub-modules: time-series indicator heatmap generation and visual semantic feature sequence extraction. (Refer to...) Figure 1 , 2 3. The specific implementation steps are as follows:

[0068] (1) Generation of time series index heatmap

[0069] The metrics cover three levels: cluster metrics, resource metrics, and task metrics. Among them:

[0070] Cluster metrics are metrics related to the Ascend Container Cloud cluster, including the number of compute resource nodes in the cluster. Number of cluster computing resource types Number of containerized task types in the cluster Number of containerized tasks in the cluster Cluster computing resource quota and the remaining rate of cluster computing resources ,in This indicates the type number of the computing resource. As one specific implementation, the computing resources of the Ascend container cloud... This encompasses the remaining utilization rates of AI cores, AI CPUs, and high-bandwidth storage, as well as the general computing resources of container cloud nodes. The AI ​​core includes matrix computation units, vector computation units, a built-in scalar computation unit, and a dedicated hardware accelerator. The AI ​​CPU includes scalar computation units and AI CPU vector computation units. The high-bandwidth storage includes bandwidth utilization and on-chip L1 / L2 cache utilization. The general computing resources of container cloud nodes are provided by Kunpeng CPUs and include CPU resources, memory, and I / O resources. These resources cover both the heterogeneous computing characteristics of the Ascend chip and supporting resources, showcasing the load differences and coordination of different resources on the node, and providing a heatmap of timing metrics for the scheduling system.

[0071] Resource metrics include scheduling metrics related to each computing resource node in the Ascend container cloud cluster, including resource quotas for the computing resource nodes. Calculate the resource remaining rate of resource nodes. Calculate the number of containerized tasks for resource nodes. and the task occupancy rate of computing resource nodes. ;in Indicates the computing resource node number. Indicates the computing resource type number, Indicates the containerization task type label. .

[0072] Time-series indicator heatmaps are used for visual representation of the Ascend container cloud environment and task execution status, consisting of a set of resource indicator heatmaps. Heatmap of task indicators Composition. Interval time. Sampling was performed using both time-continuous R-EHM and T-EHM. Zhang represents a valid timing observation in the scheduling scenario. It is represented as...

[0073]

[0074]

[0075]

[0076] in For scheduling time The time-series indicator heatmap is represented as a resource indicator heatmap sequence. Heatmap sequence of task indicators The combination; The distribution represents the length, width, and number of channels of the heatmap. For any R-EHM, the horizontal axis represents the compute resource node number. The vertical axis represents the type of computing resources. The pixel value represents the node. resource types residual rate Similarly, the horizontal axis of any T-EHM chart represents the compute resource node number. The vertical axis represents the containerized task type. The pixel value represents the task occupancy rate. , i.e., node The above type is The proportion of containerized tasks in the total number of similar tasks in the Ascend Container Cloud environment.

[0077] As a specific implementation method, sequential color mapping is used to... and Mapped to three channels pixel values, resource index heatmap pixel values and task indicator heatmap pixel values They are respectively represented as

[0078]

[0079]

[0080] in The function rounds floating-point numbers to integer values. The linear interpolation formula is as follows: Figure 3 The diagram shows how pixel values ​​in R-EHM are mapped to a grayscale heatmap, and pixel values ​​in T-EHM are mapped to an orange heatmap.

[0081] (2) Visual semantic feature sequence extraction

[0082] The visual semantic feature sequence extraction submodule uses a lightweight pre-trained ResNet-18 as the visual encoder, from... Extract the length as A high-order visual semantic feature sequence is generated. This feature sequence is linearly projected through a multilayer perceptron to align with the text embedding space of generative artificial intelligence. The multilayer perceptron comprises two sequentially connected fully connected networks, each configured with a GeLU activation layer and a LayerNorm normalization layer. This structural design ensures that the projected visual semantic feature sequence is consistent with the text embedding dimension, achieving seamless interaction and feature fusion between the visual and text modalities.

[0083] Finally, the visual semantic feature sequence processed as described above, together with the task prompts in the scheduling strategy planning module, constitutes multimodal input data, which is used to guide generative artificial intelligence to complete the corresponding scheduling strategy planning process based on the user's intent text and the feature information of the current scheduling scenario.

[0084] (ii) Scheduling Interface Engine Module

[0085] The scheduling interface engine module provides a standardized task scheduling and Ascend Container Cloud environment resource management interface library and metadirectory for the two-stage scheduling planning of environment awareness and action reasoning in the scheduling strategy planning module. It supports dynamic plug-and-play integration of existing and newly developed scheduling interfaces. The interface library ensures consistency of calling methods within the same category, providing flexible interface adaptation for scheduling strategy planning. The metadirectory accepts interface registration information, enabling unified interface registration, discovery, and invocation based on this information.

[0086] The scheduling interface engine module divides container cloud task scheduling strategy planning into two stages: scheduling environment awareness and scheduling action reasoning. It provides a standardized interface library and meta-directory for task scheduling and Ascend container cloud environment resource management. (Refer to...) Figure 1 , 4 The specific implementation steps are as follows:

[0087] (1) Scheduling environment perception

[0088] The scheduling environment-aware interface adopts a three-tier inheritance architecture, including an abstract base class, an abstract class, and an implementation class, wherein:

[0089] The abstract base class `Perception` defines the basic interface for scheduling environment awareness, mandating that subclasses implement the `cluster_metrics` and `task_metrics` methods. The `cluster_metrics` method outputs a list of structured metrics for the Ascend Container Cloud environment, and the `task_metrics` method outputs a list of structured metrics for tasks in the scheduling environment. The collection and structuring of these metrics are handled by the subclasses. These methods utilize the APIs exposed by the Ascend Container Cloud environment, combined with the specific needs of the subclasses, to achieve the acquisition and integration of cluster and task metrics.

[0090] Abstract classes inherit from the abstract base class Perception, including Vector-Based Perception (VBE), Image-Based Perception (IBE), and Graph-Based Perception (GBE). Vector-based representation predefines vector dimensions according to the scene scale, standardizing cluster indicators and tasks and mapping them to predetermined vectors. This is suitable for scheduling scenarios where task-resource interactions are clear and variable scales are controllable. Image-based representation uses visualization technology to transform the scheduling environment into a two-dimensional image. Furthermore, a pixel mapping mechanism encodes resource distribution and task locations into RGB three-channel images, enabling the class to grasp the spatial topological relationships and dynamic changes of the scheduling environment through image processing algorithms. This is suitable for scheduling scenarios with dynamically changing environmental parameters and complex spatial interactions. Graph-based representation uses nodes to represent task and container resource entities, and directed edges to represent dependencies, affinity, and anti-affinity relationships between entities. This is suitable for scheduling scenarios with deep dependencies.

[0091] Implementation classes inherit from the abstract classes VBE, IBE, or GBE, and, based on fully implementing all methods of the abstract classes, describe their own functionality, input / output types, and usage examples through commented documentation, forming a standardized and reusable interface. Finally, each interface registers its own information with the meta directory in JSON format, including the interface name, interface type, functional description, and method list.

[0092] (2) Scheduling action reasoning

[0093] The scheduling action reasoning adopts a three-tier inheritance architecture, including an abstract base class, an abstract class, and an implementation class, wherein:

[0094] The abstract base class `Inference` defines the basic interface for scheduling action reasoning, requiring subclasses to implement the scheduling decision method. The method takes the structured environment state from the scheduling environment awareness phase as input, implements the scheduling decision based on the subclass algorithm, and outputs the probability of scheduling to each node in a dictionary format.

[0095] Abstract classes inherit from the abstract base class `Inference`, including Heuristic Reasoning (HS), Swarm Intelligence-based Reasoning (SIS), Deep Learning-based Reasoning (DLS), and Deep Reinforcement Learning-based Reasoning (RLS). Heuristic Reasoning pairs tasks with resources based on deterministic rules, offering advantages such as strong interpretability and fast decision-making, making it suitable for scenarios with deterministic requirements where scheduling quality can be compromised. Swarm Intelligence-based Reasoning simulates natural life and physical processes, leveraging collaborative decision-making by numerous distributed basic units to effectively expand the search scope, making it suitable for flexible decision-making in large-scale, multi-constraint scenarios. Deep Learning-based Reasoning utilizes the predictive power of deep learning to automatically uncover hidden features in the scheduling environment to aid decision-making, suitable for scenarios with large amounts of high-quality datasets and complex scheduling rules. Deep Reinforcement Learning-based Reasoning continuously optimizes decision-making strategies through trial and error with the training environment or real scheduling clusters, suitable for scenarios with high dynamism, timely feedback, and the need for continuous adaptive adjustment.

[0096] The implementation class inherits from the abstract classes HS, SIS, DLS, or RLS, and implements the specific algorithms for scheduling decisions. These algorithms originate from three sources: first, decision algorithms migrated from open-source container scheduling platforms such as Kubernetes, Docker Swarm, and Apache Mesos, maintaining interface consistency through adaptation to suit the Ascend container cloud environment; second, drawing on the latest academic achievements, absorbing validated scheduling models from conference papers and authoritative journals to maintain the scientific rigor and technological advancement of the scheduling algorithms; and third, self-developed algorithms, innovatively designing scheduling logic to address the computing power characteristics of the Ascend chip and the collaborative scheduling needs of geospatial application computing tasks, forming a differentiated technological advantage. These algorithms constitute the main body of the implementation class, registering their information with the metadirectory, including name, type, functional description, and method list, forming an open, flexible, and pluggable interface.

[0097] (III) Scheduling Strategy Planning Module

[0098] The scheduling strategy planning module receives user intent text and visual semantic feature sequences output by the scheduling scenario analysis module as input. Relying on the container-based scheduling retrieval enhancement link and a multi-layered knowledge support system, it queries the scheduling interface engine's meta-directory to obtain the names, functional descriptions, and input / output types of all currently registered environment perception and action reasoning interfaces. These are then injected as task prompts into the API definition section, guiding the generative AI to output interface-level scheduling strategy code, which is then sent to the scheduling strategy evaluation module for verification and execution. The container-based scheduling retrieval enhancement link retrieves similar historical cases from a scheduling experience knowledge sample library, providing a reference for strategy generation. The multi-layered knowledge support system covers raw data text related to scheduling tasks and relational knowledge text generated through relational semantic extraction and reconstruction, forming a comprehensive knowledge support environment.

[0099] The scheduling strategy planning module includes sub-modules for generative artificial intelligence, task prompts, contextual learning, and knowledge base construction, which drive the scheduling interface engine to generate scheduling strategy code. (Refer to...) Figure 1 , 2 The specific implementation steps are as follows: 4, 5.

[0100] (1) Task prompt words

[0101] The task prompt word module enhances the reasoning ability of the generative AI submodule in strategy planning by organizing key scheduling information in a structured manner. A complete task prompt word template includes five core parts: API definition, system prompts, context examples, task instructions, and feedback correction instructions. Its structure is as follows: Figure 5 As shown, this aims to enhance the cognitive and reasoning abilities of generative artificial intelligence regarding tasks. Specifically:

[0102] The API definition is based on the metadirectory provided by the scheduling interface engine, injecting optional environment-aware and action-reasoning interface information. Using natural language semantic analysis, the matching degree between the structural function description text in the metadirectory and the user intent text is compared, initially filtering out interfaces with a matching degree higher than a threshold. The names, types, function descriptions, and input / output types of these interfaces are then incorporated into the API definition section of the task prompt word template.

[0103] The system prompts cover domain knowledge and Ascend Container Cloud environment information relevant to the current scheduling scenario. The domain knowledge originates from the knowledge base built within the scheduling strategy planning module, including knowledge texts of professional concepts, specification documents and industry standards, as well as best practices in typical scenarios. Furthermore, the knowledge texts of professional concepts include Ascend computing card affinity, Ascend platform network topology, and other Ascend Container Cloud environment-related concepts. The Ascend Container Cloud environment information includes cluster infrastructure configuration, Ascend Container Cloud environment monitoring indicators, and task dependencies. Domain knowledge and structured environment information encompassing user intent and environmental characteristics are incorporated into the system prompt section of the task prompt word template.

[0104] The context examples provide historical scheduling cases similar to the current scheduling scenario. These historical scheduling cases are derived from context learning in the scheduling strategy planning module and include user intent text. Resource demand vector With the preferred scheduling strategy code Based on the current task User intent text With resource demand vector A scheduling task similarity function is constructed. The historical scheduling cases with the highest similarity from the scheduling experience knowledge sample library are incorporated into the context example part of the task prompt word template.

[0105] Task instructions include a document string and a thought chain instruction. The document string summarizes the core objectives, constraints, and key execution requirements of the currently scheduled task from the user's intent text. The thought chain instruction guides a generative artificial intelligence distributed inference scheduling strategy. The document string and the thought chain instruction are incorporated into the task instruction section of the task prompt word template.

[0106] The feedback correction instructions originate from the scheduling policy evaluation module, including error feedback from the rule evaluation phase and runtime feedback from the runtime evaluation phase. Only when the scheduling policy evaluation module provides feedback correction instructions are the error feedback and runtime feedback combined into the feedback correction instruction section of the task prompt word template, allowing generative artificial intelligence to regenerate the scheduling policy code.

[0107] (2) Contextual learning

[0108] The context learning relies on a scheduling experience knowledge sample base. ,in For the first A historical scheduling task, To and The corresponding optimal scheduling strategy code, user intent text, and standardized resource requirement vector are respectively... The current task Embedding vectors are formed by embedding user intent text space into the scheduling experience knowledge sample library. Then the maximum inner product search is used. The similarity of the texts is determined. Resource similarity depends on the norm of the differences between tasks, i.e. To make it smoothly controlled in Within the range, select Characterizing resource similarity

[0109]

[0110] in Temperature is a parameter used to describe the importance of resource similarity. Sigmoid function

[0111]

[0112] The similarity function that integrates task semantics and resource matching is:

[0113]

[0114] The resource similarity function As a weighting factor for the semantic inner product similarity of the task, it effectively avoids the misselection of examples with severely deviated resource distributions. Based on the similarity function... From the scheduling experience knowledge sample base Select the one with the highest similarity Each example is used as a context example set, formally defined as:

[0115]

[0116] in Indicates the top similarity ranking An index of historical scheduling examples. Inject the contextual example portion of the task prompt word template to guide generative AI in planning effective scheduling strategies.

[0117] (3) Knowledge base construction

[0118] The knowledge base, by constructing a multi-layered knowledge support system, addresses the problems of semantic understanding bias, insufficient environmental perception, and limited reasoning ability in the professional scheduling field of generative artificial intelligence. Furthermore, the knowledge base includes an ontology sample library (OSL) and a feature sample library (FSL), wherein:

[0119] The ontology sample library stores raw data text related to scheduling tasks, specifically including task description text, infrastructure configuration, Ascend Container Cloud environment monitoring data, and domain knowledge. The task description text provides a structured description of the scheduling object, including task ID, task type, resource requests, and service target information. The infrastructure configuration provides information on the underlying resource structure of the container cloud, including node hardware specifications, cluster topology and geographical distribution, resource reachability, and availability. The Ascend Container Cloud environment monitoring data provides real-time environmental metrics, including resource utilization during task execution, node running status, task distribution, and important log information. The domain knowledge provides a basis for policy design and compliance objectives, including Ascend Container Cloud environment-related concepts, industry standards, and best practices in typical scenarios. The ontology sample library organizes data using standard JSON format and standardizes similar data from multiple heterogeneous sources through a unified naming convention.

[0120] Furthermore, the domain knowledge encompasses expertise in the Ascend container cloud architecture, focusing on the technical characteristics and best practices of the Ascend NPU and its supporting software stack. The knowledge base covers the following dimensions of expertise: First, hardware optimization knowledge, including the memory scheduling strategy of the Ascend 910 NPU and the parallel mode of multi-AI Core collaborative computing; second, operator and precision knowledge, including the principle of operator acceleration by matrix computation units in the Da Vinci architecture, applicable scenarios for mixed-precision computing, and precision compensation methods; third, deployment practice knowledge, including model conversion processes based on the CANN toolchain, scheduling adaptation of Ascend resources in a containerized environment, Ascend computing card affinity, Ascend platform network topology, and performance tuning practices and official recommendations in some typical scenarios. This knowledge forms a mapping with infrastructure configuration and cluster monitoring data in the ontology sample library, providing intelligent design for scheduling strategies oriented towards the Ascend architecture.

[0121] The feature sample library is based on the original data text of the ontology sample library. It generates a set of relational knowledge texts through relational semantic extraction and reconstruction, specifically including dependency relation knowledge texts, affinity relation knowledge texts, and anti-affinity relation knowledge texts. The dependency relation knowledge texts describe the logical constraints on the execution order between tasks, reflecting the temporal relationships and data flow dependencies between tasks. The affinity relation texts describe the compatibility relationships between tasks and between tasks and containers, reflecting a tendency for centralized task deployment or specific requirements for containers. The anti-affinity relation texts describe the exclusion relationships between tasks and between tasks and containers, reflecting a tendency for distributed task deployment. The feature sample library is based on the ontology sample library.<subject,relation, object> The triplet structured representation describes the relationship between instance `subject` and instance `object`. The instance `subject` and instance `object` share an indexing mechanism with the ontology sample library, supporting efficient association retrieval of raw data text and extracted feature data text.

[0122] (iv) Scheduling strategy evaluation module

[0123] The scheduling strategy evaluation module is used to verify the execution of generative scheduling strategies. It constructs a reflective mechanism encompassing rule evaluation and runtime evaluation to enable timely correction of the input strategy. Feedback information is then passed to the scheduling strategy planning module, triggering a new round of strategy planning and achieving a positive feedback loop of strategy reflection and correction, thus mitigating the risks of invalid or backward reasoning. Simultaneously, it outputs and archives verified high-quality scheduling strategies, forming a scheduling experience knowledge sample library. This library allows the scheduling strategy planning module to proactively retrieve and reference these strategies during future task execution, maintaining a positive feedback loop of strategy reflection and correction.

[0124] The scheduling strategy evaluation includes rule evaluation (Evaluator-T1) and runtime evaluation (Evaluator-T2) mechanisms, used to provide feedback on inference results and operational performance. Through generative artificial intelligence self-reflection and strategy correction, it avoids the risks of invalid or reverse inference. (Refer to...) Figure 1 , 2 5. The specific implementation steps are as follows:

[0125] (1) Rule evaluation

[0126] The rule evaluation mechanism provides feedback on the inference results. The scheduling strategy output is statically validated through interface verification rules, affinity verification rules, and conflict verification rules. Wherein:

[0127] Interface validation rules are used to ensure that the data format and logic between the scheduling environment awareness interface and the scheduling action reasoning interface in the scheduling strategy generated by generative artificial intelligence are consistent, thereby ensuring that the strategy combination can be executed reasonably. Furthermore, for any scheduling strategy combination... ,in For scheduling environment-aware interfaces, For the scheduling action reasoning interface, verification

[0128]

[0129] in for Input type and Output type matching ensures the correct data structure is passed. For environmental perception interface Under the premise of semantic output, the action reasoning interface The existing contextual constraints are satisfied, ensuring the semantic rationality of interface composition.

[0130] Affinity verification rules are used to ensure that the scheduling strategies generated by generative AI conform to node affinity and task affinity constraints. Let the scheduling mapping be... ,in Represents a set of scheduled tasks. This represents a set of nodes, where different scheduled tasks are allowed to be deployed to the same node. The node affinity constraint in FSL can be defined by... It means that, among them The verification rules are as follows: The task affinity constraint in FSL can be determined by... It means that, among them The verification rules are as follows: .

[0131] Conflict checking rules are used to ensure that scheduling strategies generated by generative AI do not introduce unauthorized resource contention or exclusive conflicts. In FSL, conflict constraints can be defined by... This indicates that the verification rule is... .

[0132] When the interface validation rules, affinity validation rules, and conflict validation rules all pass, the interface-level scheduling strategy code can be issued and executed; when the validation rules fail, the violation situation is structured into error feedback information and sent back to the scheduling strategy planning module to regenerate the strategy.

[0133] (2) Runtime evaluation

[0134] The runtime evaluation mechanism provides feedback on the performance of running tasks. Through task-level error tolerance and cluster-level target tolerance, it ensures that the personalized requirements of heterogeneous scheduling tasks are met. The task-level error tolerance is set as follows: The mission-level objective is ,in To schedule tasks, For environmental nodes, The representative will Deploy to The extent to which personalized requirements for the task are violated. This applies to latency-sensitive tasks. , This manifests as a standardized response time; for cost-driven tasks, This manifests as standardized scheduling overhead. The cluster-level error tolerance is set to... Cluster-level target is This is reflected in the degree of resource balance, i.e., the standard deviation of resource utilization distribution among containers. The runtime evaluation function is...

[0135]

[0136] in For scheduling nodes Remaining resources For scheduling tasks The amount of resources requested. When When the value is False, the violation is structured into runtime feedback information and sent back to the scheduling strategy planning module to regenerate the strategy; when the verification passes, the user intent text, resource request vector, and scheduling decision code are structured into standard JSON format and archived into the scheduling experience knowledge sample library.

[0137] (v) Data Augmentation Module

[0138] The data augmentation module enhances text and image data to form an expanded sample set. It interfaces with the scheduling experience knowledge sample library output and archived by the scheduling strategy evaluation module, augmenting the user intent text and time-series indicator heatmaps contained in the scheduling experience knowledge samples, and constructing image-text sample pairs with high-quality scheduling strategies included in the scheduling experience knowledge samples. The output image-text sample pairs are used to continuously fine-tune the generative artificial intelligence in the scheduling strategy planning module to improve the generative AI's anti-interference ability and generalization ability. The text data augmentation strategies include synonym replacement, noise injection, and semantic invariant rewriting; the image data augmentation strategies include random cropping of time-series indicator heatmaps, noise injection, and random occlusion; the image-text sample pairs are constructed based on positive and negative samples composed of the original data and the augmented data.

[0139] The data augmentation module includes text data augmentation, image data augmentation, and image-text pair construction, used to train generative artificial intelligence in scheduling strategy planning, improving its anti-interference ability and generalization ability. (Refer to...) Figure 1 , 6 The specific implementation steps are as follows:

[0140] (1) Text data augmentation

[0141] The text data augmentation strategy utilizes synonym replacement, noise injection, and semantic invariance rewriting to generate enhanced text data for training generative artificial intelligence. The synonym replacement strategy randomly replaces random terms 'u' in the user's intent text with terms 'v' from the thesaurus, improving the system's generalization ability to diverse terms. The noise injection strategy randomly inserts noise into the user's intent text, including irrelevant words, typos, punctuation marks, and out-of-domain terms, simulating interference in real input to improve system robustness. The semantic invariance rewriting strategy performs sentence transformation and word order adjustment on the user's intent text, generating diverse expressions while maintaining semantic invariance, enhancing the system's ability to understand semantically equivalent expressions.

[0142] (2) Image data enhancement

[0143] The image data enhancement employs random cropping, noise injection, and random occlusion strategies to expand and generate a large number of time-series indicator heatmap sample data, overcoming the challenge of sparsity in production-grade samples. The random cropping strategy is applied to the time-series indicator heatmap... Select different regions Cut the samples to create different sizes. To simulate container cloud environments of varying sizes and task scheduling lengths, the noise injection strategy injects different levels of Gaussian noise into the time-series metric heatmap. Generate noisy samples This simulates random interference caused by network fluctuations and resource contention under different scenarios. The random occlusion strategy is used to simulate random loss areas in the time-series index heatmap. Information, generating occlusion samples This is to simulate resource unavailability and unavailability caused by unforeseen events.

[0144] (3) Image-text pair construction

[0145] The image-text pair construction is based on augmented data to form positive / negative sample pairs, supporting generative artificial intelligence in scheduling decisions under difficult scenarios, including those based on raw image data. Enhanced image data obtained by enhancing the image data. Constructing image sample pairs, and from the original image data User intent text Enhanced text obtained by enhancing the text data. , with the optimal scheduling strategy code Construct image-text sample pairs. Where:

[0146] Image sample pairs are derived from a set of time-series index heatmaps and their respective enhanced images. Each sample is anchored from the original image data. Its corresponding enhanced image data Constitute positive sample pairs , and the remainder Individual and The non-corresponding original image data and enhanced image data constitute negative sample pairs, i.e.

[0147]

[0148] Image-text sample pairs are derived from a set of time-series indicator heatmaps. One sample, user intent text and its enhanced text and scheduling strategy code The corresponding time-series indicator heatmap and text data form positive sample pairs, i.e.

[0149]

[0150] Remaining Zhang Fei's corresponding time-series indicator heatmap and text data constitute a negative sample pair, that is...

[0151]

[0152] (vi) Ascend Container Cloud Cluster

[0153] The Ascend Container Cloud Cluster provides resource support for geospatial container applications, including the domestically produced Ascend NPU chip specifically designed for artificial intelligence tasks and the general-purpose support resource Kunpeng CPU.

[0154] The Ascend container cloud cluster uses the Ascend NPU chip, designed specifically for artificial intelligence tasks, as its core computing power support, coupled with the general-purpose Kunpeng CPU, to build a high-performance, domestically produced hardware foundation, providing comprehensive resource support for various container applications. These container applications cover multi-source, heterogeneous geospatial application models with cross-access, as well as inference and training tasks for large-scale geographic information models. Specifically:

[0155] The Ascend NPU chip, a dedicated chip for artificial intelligence computing, is architecturally structured around efficiently processing AI tasks. It includes an AI computing core, intelligent storage, a heterogeneous computing interface, and a low-power design unit. The AI ​​computing core encompasses matrix computing units, vector computing units, a built-in scalar computing unit, and a dedicated hardware accelerator, supporting forward inference and backward propagation of complex neural networks. The intelligent storage, through a multi-level caching architecture and high-bandwidth storage interfaces, enables rapid data transfer and adapts to real-time interaction with massive amounts of data and models. The heterogeneous computing interface communicates efficiently with other processors such as CPUs and GPUs, supporting collaborative computing power across different scenarios. The low-power design unit optimizes energy efficiency while meeting computing power requirements through dynamic voltage adjustment and refined power control. Considering the characteristics of the Ascend container cloud environment, the scheduling scenario analysis module utilizes Ascend container cloud resources to construct a heatmap of time-series indicators; the scheduling strategy planning module incorporates the domain knowledge of the Ascend NPU chip into the system prompt section of the task prompt word template.

[0156] Based on differences in resource-oriented types, geospatial application models can be categorized into CPU-intensive, memory-intensive, disk-intensive, bandwidth-intensive, and resource-inefficient models. Specifically: CPU-intensive models primarily include image processing and spatial computing tasks such as intelligent interpretation of remote sensing images and vector topology analysis; memory-intensive models encompass data extraction and product production tasks such as multi-source geographic data fusion and extraction and automated production of series maps; disk-intensive models cover offline (batch) processing tasks such as historical geographic data archiving and processing and full-domain terrain modeling; bandwidth-intensive models focus on data access tasks such as distributed spatial database access and real-time remote sensing data reception; and resource-inefficient models mainly include lightweight computing tasks such as basic spatial query analysis, regional ecological environment assessment, and urban and rural planning decision support. For geospatial application models on the Ascend Container Cloud environment, the scheduling strategy evaluation module saves high-quality scheduling results to a scheduling experience knowledge sample library, and the scheduling strategy planning module actively accesses this library to assist in planning the scheduling of various geospatial application models on the Ascend Container Cloud environment. Inference training of large-scale geographic information models is a special type of container cloud application task, exhibiting different resource requirements than ordinary container cloud application tasks. The large-scale model training phase places extremely high demands on Ascend NPU computing power, memory capacity, and storage bandwidth, and requires prolonged use of large-scale cluster resources. The large-scale model inference phase emphasizes response timeliness and concurrent processing capabilities, with resource requirements dynamically fluctuating with the number of query requests, placing higher demands on the elastic scaling and load balancing capabilities of the container cluster. For the inference training task of large-scale geographic information models on the Ascend container cloud environment, the scheduling strategy evaluation module saves high-quality scheduling results to a scheduling experience knowledge sample library, and the scheduling strategy planning module actively accesses the scheduling experience knowledge sample library and combines it with the domain knowledge of the Ascend container cloud environment within the knowledge library to plan the scheduling.

[0157] Based on the aforementioned end-to-end generative scheduling system, this invention also proposes an end-to-end generative scheduling method for the Ascend container cloud environment, specifically including the following steps:

[0158] By using the scheduling scenario analysis module, time-series correlated indicator data is extracted from the Ascend Container Cloud environment as input. Through the aggregation and visualization of the time-series indicators, a heatmap of time-series indicators of the current resource usage status and task distribution of the Ascend Container Cloud environment is output to the generative scheduling system. This heatmap is then transformed into a visual semantic feature sequence to enhance the model's understanding of the scheduling environment.

[0159] By utilizing the scheduling interface engine module, a standardized task scheduling and Ascend container cloud environment resource management interface library and metadirectory are provided for the two-stage scheduling planning of environment perception and action reasoning in the scheduling strategy planning module, supporting dynamic plug-and-play integration of existing scheduling and newly developed scheduling interfaces.

[0160] The scheduling strategy planning module receives user intent text and visual semantic feature sequences output by the scheduling scenario analysis module as input. Relying on the container scheduling retrieval enhancement link and multi-level knowledge support system, it guides the generative artificial intelligence to output interface-level scheduling strategy code.

[0161] The scheduling strategy evaluation module verifies the generative scheduling strategy. A reflective mechanism encompassing rule evaluation and runtime evaluation is constructed to enable timely correction of the input strategy. Feedback information is then transmitted to the scheduling strategy planning module, triggering a new round of strategy planning and achieving a positive feedback loop of strategy reflection and correction, thus mitigating the risks of invalid or reverse reasoning. Simultaneously, high-quality, verified scheduling strategies are output and archived, forming a scheduling experience knowledge sample library. This library allows the scheduling strategy planning module to proactively retrieve and reference these strategies during future task execution, maintaining a positive feedback loop of strategy reflection and correction.

[0162] By utilizing the data augmentation module and connecting to the scheduling experience knowledge sample library of the scheduling strategy evaluation module, text and image data are augmented to form an expanded sample set, and image-text sample pairs are constructed. This allows for continuous fine-tuning of the generative artificial intelligence in the scheduling strategy planning module to improve its anti-interference and generalization capabilities. The text data augmentation strategies include synonym replacement, noise injection, and semantic invariant rewriting. The image data augmentation strategies include random cropping of time-series indicator heatmaps, noise injection, and random occlusion. The image-text sample pairs are constructed based on the original data and augmented data to form positive and negative samples.

[0163] Based on the Ascend container cloud cluster, including the domestically produced Ascend NPU chip specifically designed for artificial intelligence tasks and the general-purpose support resource Kunpeng CPU, resource support is provided for container applications. These container applications include multi-source heterogeneous, cross-access geospatial application models, as well as inference training of large-scale geographic information models. Based on the different types of resource preferences, the geospatial application models include CPU-intensive, memory-intensive, disk-intensive, bandwidth-intensive, and resource-insensitive models. Furthermore, the inference training of the large-scale geographic information models exhibits significant characteristics in terms of resource requirements: the training phase has high requirements for Ascend NPU computing power and memory capacity, occupying cluster resources for a long time, making it a typical computing-intensive task; the inference phase emphasizes timeliness and concurrent processing capabilities, with resource requirements fluctuating significantly with the number of requests, placing higher demands on the load balancing capabilities of the container cloud environment.

[0164] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements an end-to-end generative scheduling method for the Ascend container cloud environment, thereby achieving end-to-end generative scheduling for the Ascend container cloud environment.

[0165] A computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, the end-to-end generative scheduling method for the Ascend container cloud environment is implemented, thereby realizing end-to-end generative scheduling for the Ascend container cloud environment.

[0166] In summary, the end-to-end generative scheduling system for the Ascend container cloud environment achieves an end-to-end closed-loop solution for environment awareness, action reasoning, and policy evaluation feedback in containerized distributed clusters through scheduling scenario analysis, scheduling interface engine, scheduling policy planning, and scheduling policy evaluation. This improves the flexibility and stability of scheduling decisions in the Ascend container cloud environment and gives the system the adaptive capability to dynamically adjust policies as the environment changes.

[0167] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0168] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An end-to-end generative scheduling system for the Ascend container cloud environment, characterized in that, It includes, in sequence, a scheduling scenario analysis module, a scheduling strategy planning module, a scheduling strategy evaluation module, a scheduling interface engine module, and a data enhancement module, wherein: The scheduling scenario analysis module is used to extract resource indicators and task indicators with time-series correlation from the Ascend container cloud environment, generate a time-series indicator heatmap composed of resource indicator heatmap sequences and task indicator heatmap sequences, and extract visual semantic feature sequences from the time-series indicator heatmap and output them to the scheduling strategy planning module. The scheduling strategy planning module is used to receive user intent text and visual semantic feature sequences, query the meta directory of the scheduling interface engine module, retrieve similar historical cases from the scheduling experience knowledge sample library by fusing task semantics and resource matching similarity function, and construct structured task prompt words by combining domain knowledge obtained from multi-level knowledge support system, so as to guide the generative artificial intelligence to output executable interface-level scheduling strategy code, and output it to the scheduling strategy evaluation module. The scheduling strategy evaluation module is used to perform rule evaluation and runtime evaluation on the received interface-level scheduling strategy code, and send the error feedback or runtime feedback information corresponding to the strategy that fails the evaluation back to the scheduling strategy planning module to regenerate the strategy. At the same time, the scheduling strategy code that passes the evaluation is archived to the scheduling experience knowledge sample library. The scheduling interface engine module is used to provide a standardized task scheduling and Ascend container cloud environment resource management interface library and meta directory for the scheduling strategy planning module. The interface library includes a scheduling environment awareness interface and a scheduling action reasoning interface. The meta directory includes interface name, interface type, function description and method list. The data augmentation module is used to process the original data through text data augmentation strategies and image data augmentation strategies to generate an expanded sample set and image-text sample pairs to train the generative artificial intelligence in the scheduling strategy planning module.

2. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, The scheduling scenario analysis module includes a time-series indicator heatmap generation submodule and a visual semantic feature sequence extraction submodule, wherein: The time-series indicator heatmap generation submodule is used to generate a time-series indicator heatmap composed of a resource indicator heatmap sequence and a task indicator heatmap sequence. The specific method is as follows: Based on the resource remaining rate of computing resource nodes The first linear mapping formula is used to convert the values ​​into grayscale values, generating a series of resource indicator heatmaps. ； Task occupancy rate based on computing resource nodes The values ​​are converted into orange series values ​​through the second linear mapping formula, generating a heatmap sequence of task indicators. ； in, The function rounds a floating-point number to an integer value. Indicates the computing resource node number. Indicates the computing resource type number, Indicates the containerized task type label; The visual semantic feature sequence extraction submodule is used to extract visual semantic feature sequences, and the specific method is as follows: A pre-trained visual encoder is used to extract high-order visual features from a time-series index heatmap; By using a multilayer perceptron consisting of two fully connected networks connected in sequence, each configured with a GeLU activation layer and a LayerNorm normalization layer, linear projection is performed to align high-order visual features to the text embedding space of generative artificial intelligence, forming a sequence of visual semantic features.

3. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, In the scheduling interface engine module Both the scheduling environment awareness interface and the scheduling action reasoning interface adopt a three-tier inheritance architecture including an abstract base class, an abstract class, and an implementation class. The abstract classes of the scheduling environment awareness interface include a perception class based on vector representation, a perception class based on image representation, and a perception class based on graph representation. The abstract classes of the scheduling action reasoning interface include heuristic reasoning classes, swarm intelligence-based reasoning classes, deep learning-based reasoning classes, and deep reinforcement learning-based reasoning classes.

4. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, The scheduling strategy planning module constructs task prompts, including: The API definition section uses semantic analysis to compare the user intent text with the functional description text of the interface in the meta directory of the scheduling interface engine module, and filters out interface information with a matching degree higher than a preset threshold. The system prompts are based on a multi-layered knowledge support system, including Ascend Container Cloud architecture expertise, infrastructure configuration information, Ascend Container Cloud environment monitoring data, and task dependencies. The context example section calculates the semantic similarity of tasks and the similarity of resource requirements, and retrieves the most similar historical scheduling cases from the scheduling experience knowledge sample library. ； in, For the current task, For the l-th historical scheduling task, for User intent text embedding vector, for User intent text embedding vector, The formula for calculating resource similarity based on resource demand vectors is as follows: ； in Temperature is a parameter used to describe the importance of resource similarity. For the Sigmoid function, and These are the standardized resource requirement vectors for the current task and the historically scheduled tasks, respectively. The task instruction section includes a document string summarized from the user intent text and a thought chain instruction that guides distributed reasoning. The document string summarizes the core objectives, constraints, and key execution requirements of the current scheduled task from the user intent text. In the feedback correction instruction section, when feedback information is received from the scheduling strategy evaluation module, the feedback information and the original scheduling strategy code are incorporated into the task prompt word.

5. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, The scheduling policy evaluation module performs rule evaluation and runtime evaluation, wherein: The rule evaluation includes interface validation rule evaluation, affinity validation rule evaluation, and conflict validation rule evaluation, wherein: Interface validation rule evaluation is used to verify whether the output data type of the scheduling environment awareness interface matches the input data type of the scheduling action reasoning interface in the interface-level scheduling strategy code, and whether the semantic constraints of the interface combination are met. Affinity verification rule evaluation is used to verify whether the scheduling mapping relationship defined by the interface-level scheduling strategy code conforms to the node affinity constraint and the task affinity constraint; The conflict verification rule evaluation is used to verify whether the scheduling mapping relationship defined by the interface-level scheduling policy code avoids resource contention or exclusive conflicts. The runtime evaluation includes task-level error tolerance evaluation and cluster-level error tolerance evaluation, wherein: Task-level error tolerance assessment uses standardized response time as the indicator for latency-sensitive tasks and standardized scheduling overhead as the indicator for cost-driven tasks. Cluster-level error tolerance assessment uses the standard deviation of resource utilization distribution among containers as an indicator.

6. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, In the data enhancement module The text data augmentation strategies include synonym replacement, noise injection, and semantic invariant rewriting of user intent text. The image data enhancement strategy includes random cropping, noise injection, and random occlusion of the time-series indicator heatmap. The image-text sample pairs include positive sample pairs and negative sample pairs; the positive sample pairs consist of the original time-series indicator heatmap and its corresponding enhanced image, user intent text and its enhanced text, and scheduling strategy code; the negative sample pairs consist of unrelated time-series indicator heatmaps and text data.

7. The end-to-end generative scheduling system for the Ascend container cloud environment according to claim 1, characterized in that, The system is deployed on the Ascend Container Cloud cluster, which includes Ascend NPU and Kunpeng CPU. The container applications running in the Ascend Container Cloud environment include inference and training tasks for geospatial application models and large geographic information models. Based on the differences in resource preference types, the geospatial application model is divided into CPU-intensive, memory-intensive, disk-intensive, bandwidth-intensive, and resource-inefficient types.

8. An end-to-end generative scheduling method for the Ascend container cloud environment, characterized in that, An end-to-end generative scheduling system for the Ascend container cloud environment according to any one of claims 1 to 7, the method comprising: By using the scheduling scenario analysis module, time-series correlation indicators are extracted from the Ascend container cloud environment, a time-series indicator heatmap is generated, and a visual semantic feature sequence is extracted. The scheduling strategy planning module receives user intent text and visual semantic feature sequences, queries the metadirectory of the scheduling interface engine module, and guides generative artificial intelligence to generate interface-level scheduling strategy code. The scheduling strategy evaluation module performs rule evaluation and runtime evaluation on the interface-level scheduling strategy code. If the evaluation fails, feedback information is generated and sent back to the scheduling strategy planning module to regenerate the strategy. If the evaluation passes, the scheduling strategy code is archived and deployed for execution. Through the scheduling interface engine module, a standardized task scheduling and Ascend Container Cloud environment cluster resource management interface library and meta-directory are provided for the two-stage scheduling planning of environment perception and action reasoning in the scheduling strategy planning module, supporting dynamic plug-and-play integration of existing scheduling and newly developed scheduling interfaces. The data augmentation module enhances the original text and image data, generating an expanded sample set and image-text sample pairs for training generative artificial intelligence.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the end-to-end generative scheduling method for the Ascend container cloud environment as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the end-to-end generative scheduling method for the Ascend container cloud environment as described in claim 8.

Images