CN-121579226-B - End-to-end generation type dispatching system and method for rising container cloud environment

CN121579226BCN 121579226 BCN121579226 BCN 121579226BCN-121579226-B

Abstract

The invention discloses an end-to-end generation type scheduling system and a method thereof for a lifting container cloud environment, wherein a scheduling scene analysis module generates a time sequence index thermodynamic diagram and extracts visual semantic feature sequences, a scheduling strategy planning module receives user intention texts and the visual semantic feature sequences, queries meta-directories of a scheduling interface engine module, retrieves similar historical cases from a scheduling experience knowledge sample base to construct a structured task prompt word, guides a generation type artificial intelligence to output executable interface-level scheduling strategy codes, a scheduling strategy evaluation module carries out rule evaluation and operation time evaluation on the interface-level scheduling strategy codes, the scheduling interface engine module provides standardized task scheduling and lifting container cloud environment resource management interface base and meta-directories, and a data enhancement module generates an expansion sample set and image-text sample pair to train the generation type artificial intelligence in the scheduling strategy planning module. The invention improves the flexibility and stability of the scheduling decision of the cloud environment of the rising container.

Inventors

ZHU LILU
HUANG KAI
Su Yingze

Assignees

苏州空天信息研究院

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. The end-to-end generation type scheduling system for the lifting container cloud environment is characterized by comprising a scheduling scene analysis module, a scheduling strategy planning module, a scheduling strategy evaluation module, a scheduling interface engine module and a data enhancement module which are sequentially connected, wherein: The scheduling scene analysis module is used for extracting resource indexes and task indexes with time sequence relevance from a rising container cloud environment, generating a time sequence index thermodynamic diagram composed of a resource index thermodynamic diagram sequence and a task index thermodynamic diagram sequence, extracting a visual semantic feature sequence from the time sequence index thermodynamic diagram, and outputting the visual semantic feature sequence to the scheduling strategy planning module; The scheduling strategy planning module is used for receiving user intention texts and visual semantic feature sequences, inquiring a meta-catalog of the scheduling interface engine module, searching similar historical cases from a scheduling experience knowledge sample base through a similarity function integrating task semantics and resource matching, constructing a structured task prompt word by combining domain knowledge acquired from a multi-level knowledge support system, guiding a generated artificial intelligence to output an executable interface-level scheduling strategy code, and outputting the executable interface-level scheduling strategy code to the scheduling strategy evaluation module; The scheduling policy evaluation module is used for executing rule evaluation and runtime evaluation on the received interface-level scheduling policy codes, sending error feedback or runtime feedback information corresponding to the policy which is not passed by the evaluation back to the scheduling policy planning module to regenerate the policy, and archiving the scheduling policy codes which are passed by the evaluation to the scheduling experience knowledge sample library; The scheduling interface engine module is used for providing a standardized task scheduling and lifting container cloud environment resource management interface library and a meta-directory for the scheduling strategy planning module, wherein the interface library comprises a scheduling environment perception interface and a scheduling action reasoning interface, and the meta-directory comprises an interface name, an interface type, a function description and a method list; The data enhancement module is used for processing the original data through a text data enhancement strategy and an image data enhancement strategy to generate an expansion sample set and an image-text sample pair so as to train the generation type artificial intelligence in the scheduling strategy planning module.
2. The end-to-end generation scheduling system for a lifting container cloud environment of claim 1, wherein the scheduling scene analysis module comprises a time sequence index thermodynamic diagram generation sub-module and a visual semantic feature sequence extraction sub-module, wherein: the time sequence index thermodynamic diagram generating submodule is used for generating a time sequence index thermodynamic diagram composed of a resource index thermodynamic diagram sequence and a task index thermodynamic diagram sequence, and the specific method is as follows: Resource residual rate based on computing resource node Converting the first linear mapping formula into a gray value to generate a resource index thermodynamic diagram sequence; ; Task occupancy based on computing resource nodes Converting the first linear mapping formula into an orange color system value to generate a task index thermodynamic diagram sequence; ; Wherein, the The function rounds floating point numbers to integer values, Representing the number of the node of the computing resource, Representing the number of the type of computing resource, A containerized task type label; the visual semantic feature sequence extraction submodule is used for extracting visual semantic feature sequences, and the specific method comprises the following steps: extracting high-order visual features from the time sequence index thermodynamic diagram by adopting a pre-trained visual encoder; and (3) performing linear projection by a multi-layer perceptron comprising two fully connected networks which are sequentially connected and are respectively provided with GeLU activating layers and LayerNorm normalizing layers, and aligning high-order visual features to a text embedding space of the generated artificial intelligence to form a visual semantic feature sequence.
3. The end-to-end generation scheduling system for a lifting container cloud environment of claim 1, wherein, in said scheduling interface engine module, The scheduling environment perception interface and the scheduling action reasoning interface both adopt a three-layer inheritance system structure comprising an abstract base class, an abstract class and an implementation class; the abstract class of the scheduling environment perception interface comprises a perception class based on vector characterization, a perception class based on image characterization and a perception class based on graph characterization; The abstract class of the scheduling action reasoning interface comprises a heuristic reasoning class, a group-based intelligent reasoning class, a deep learning-based reasoning class and a deep reinforcement learning-based reasoning class.
4. The end-to-end generation scheduling system for a lifting container cloud environment of claim 1, wherein the scheduling policy programming module constructs task prompt words comprising: the API definition part is used for comparing the user intention text with the function description text of the interface in the metadirectory of the scheduling interface engine module through semantic analysis, and screening out interface information with the matching degree higher than a preset threshold value; The system prompt part is derived from a multi-level knowledge support system and comprises a rising container cloud architecture expertise, infrastructure configuration information, rising container cloud environment monitoring data and task dependency relations; the context example part is used for obtaining a historical scheduling case with highest similarity from a scheduling experience knowledge sample base by calculating task semantic similarity and resource demand similarity; ; Wherein, the For the present task it is possible that, The task is scheduled for the first history, Is that Is used to embed the vector in the user's intended text, Is that Is used to embed the vector in the user's intended text, For the resource similarity based on the resource demand vector, the calculation formula is: ; Wherein the method comprises the steps of Describing the importance of the resource similarity for the temperature parameter, As a function of the Sigmoid, And Standardized resource demand vectors for the current task and the historical scheduling task respectively; A task instruction part including a document character string summarized from the user intention text and a thought chain instruction guiding distributed reasoning, wherein the document character string summarizes the core target, constraint condition and key execution requirement of the current dispatching task from the user intention text; And the feedback correction instruction part is used for incorporating the feedback information and the original scheduling strategy code into the task prompt word when receiving the feedback information of the scheduling strategy evaluation module.
5. The end-to-end generation scheduling system of claim 1, wherein the scheduling policy evaluation module performs rule evaluation and runtime evaluation, wherein: The rule evaluation, interface check rule evaluation, affinity check rule evaluation, and conflict check rule evaluation, wherein: the interface verification rule evaluation is used for verifying whether the output data type of the scheduling environment awareness interface is matched with the input data type of the scheduling action reasoning interface or not and whether the semantic constraint condition of interface combination is met or not in the interface-level scheduling policy code; the affinity checking rule evaluation is used for checking whether the scheduling mapping relation defined by the interface level scheduling policy code accords with the node affinity constraint and the task affinity constraint; A conflict check rule evaluation for checking whether the scheduling mapping relation defined by the interface-level scheduling policy code avoids resource contention or exclusive conflict; the runtime evaluation includes a task level error margin evaluation and a cluster level error margin evaluation, wherein: task level error margin assessment takes standardized response time as an index for delay-sensitive tasks and standardized scheduling overhead as an index for cost-driven tasks; Cluster level error tolerance assessment is indexed by the standard deviation of resource utilization distribution among containers.
6. An end-to-end generation scheduling system for a lifting container cloud environment as recited in claim 1, wherein, in said data enhancement module, The text data enhancement strategy comprises synonym replacement, noise injection and semantic invariance rewriting of a user intention text; the image data enhancement strategy comprises the steps of carrying out random clipping, noise injection and random shielding on a time sequence index thermodynamic diagram; the image-text sample pair comprises a positive sample pair and a negative sample pair, wherein the positive sample pair is composed of an enhanced image corresponding to the original time sequence index thermodynamic diagram, a user intention text and an enhanced text thereof and a scheduling policy code, and the negative sample pair is composed of an uncorrelated time sequence index thermodynamic diagram and text data.
7. An end-to-end generation scheduling system for a lifting container cloud environment as recited in claim 1, wherein said system is deployed in a lifting container cloud cluster comprising a lifting NPU and a spread CPU, said container applications running in the lifting container cloud environment comprising a geospatial application model and a geographic information big model's inference training tasks; The geospatial application model is divided into CPU intensive, memory intensive, disk intensive, bandwidth intensive and resource non-intensive according to the distinction of the resource trend types.
8. An end-to-end generation type scheduling method for a lifting container cloud environment, which is characterized in that the end-to-end generation type scheduling system for the lifting container cloud environment according to any one of claims 1 to 7 comprises the following steps: Extracting time sequence association indexes from a rising container cloud environment through a scheduling scene analysis module, generating a time sequence index thermodynamic diagram and extracting a visual semantic feature sequence; Receiving user intention text and visual semantic feature sequences through a scheduling strategy planning module, inquiring a metadirectory of a scheduling interface engine module, and guiding a generating type artificial intelligence to generate an interface-level scheduling strategy code; The method comprises the steps of carrying out rule evaluation and operation time evaluation on a dispatching strategy code of a port level through a dispatching strategy evaluation module, generating feedback information and sending back to a dispatching strategy planning module to regenerate a strategy if the evaluation is not passed, archiving the dispatching strategy code and issuing and executing if the evaluation is passed; The scheduling interface engine module is used for providing a standardized task scheduling and lifting container cloud environment cluster resource management interface library and metadirectory for the environment awareness and action reasoning two-stage scheduling planning in the scheduling strategy planning module, and supporting the dynamic plug-in integration of the existing scheduling and newly-developed scheduling interfaces; And the data enhancement module is used for enhancing the original text data and the image data to generate an expansion sample set and an image-text sample pair for training the generated artificial intelligence.
9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the end-to-end generation scheduling method for a lifting container cloud environment of claim 8 when the computer program is executed by the processor.
10. A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a peer-to-peer generation scheduling method for a lifting container cloud environment as recited in claim 8.

Description

End-to-end generation type dispatching system and method for rising container cloud environment Technical Field The invention relates to the technical field of container cloud task scheduling and artificial intelligence intersection, in particular to an end-to-end generation type scheduling system and method for a rising container cloud environment. Background The container cloud becomes a mainstream infrastructure for supporting geospatial application models and geographic information large model training reasoning and collaborative services by virtue of the characteristics of light weight, high elasticity, agility deployment and the like. The task scheduling of the container is used as an important support for cloud computing resource management, and the core aim is to optimize the dynamic matching of the task instance and the container according to the user demands, the task characteristics and the real-time state of the cluster resources. The current mainstream method depends on static configuration and manual templates, so that the maintenance cost is high, and the current mainstream method is more difficult to adapt to dynamic requirements such as service coordination, cross access, real-time interaction and the like among geographic models. In addition, most systems lack an effective scheduling index evaluation and feedback mechanism, and scheduling imbalance is easy to occur in large-scale heterogeneous clusters in which a geospatial application model and a geographic information large model are mixed and deployed. In order to realize the complete autonomous control of computing power resources, the construction of a container cloud environment based on a rising base becomes the first choice in the industry. However, unlike the scheduling of general computing power such as CPU, GPU resources, the scheduling of container tasks for a rising container cloud environment faces greater challenges. Because the lifting-series NPU is based on the Davinci architecture, a computing unit and a hardware accelerator which are different from the traditional processor are designed, on one hand, a dispatching strategy based on general platform optimization is difficult to be directly applicable, and on the other hand, higher requirements are put forward for fine-grained management of heterogeneous resources and effective achievement of diversified service level targets. In recent years, the artificial intelligence technology provides a new path for breaking through the bottleneck of the traditional scheduling technology by virtue of autonomous decision and knowledge learning capability. In particular, the multi-modal transformation and knowledge generation capabilities of the generated artificial intelligence enable the preference of dynamic scheduling strategies in a rising container cloud environment. However, the current application still faces a great challenge, namely, the lack of a domain knowledge system deeply fused with the cloud environment of the rising container leads to deviation between the generation strategy and the actual demand, the generation result is mostly decision advice, the generation result needs to be manually developed into an executable code for the second time, and the generation result has risks of logic conflict, demand violation and the like due to the lack of an effective verification and feedback mechanism. In summary, in the field of container cloud scheduling, especially in the cloud computing environment of the domestic rising container, there is still a lack of a highly reliable end-to-end generation type scheduling solution covering "intention understanding-policy generation-result verification-feedback optimization". Disclosure of Invention The invention aims to provide an end-to-end generation type scheduling system and method for a rising container cloud environment, so as to enhance understanding of scheduling intention and dynamic adaptability to the rising container cloud environment and improve scheduling performance of container tasks in a domestic environment. The technical scheme of the invention is that the end-to-end generation type scheduling system facing the rising container cloud environment comprises a scheduling scene analysis module, a scheduling strategy planning module, a scheduling strategy evaluation module, a scheduling interface engine module and a data enhancement module which are connected in sequence, wherein: The scheduling scene analysis module is used for extracting resource indexes and task indexes with time sequence relevance from a rising container cloud environment, generating a time sequence index thermodynamic diagram composed of a resource index thermodynamic diagram sequence and a task index thermodynamic diagram sequence, extracting a visual semantic feature sequence from the time sequence index thermodynamic diagram, and outputting the visual semantic feature sequence to the scheduling strategy planning module; The scheduling strategy p