A semantically enhanced scheduling decision system and method for geospatial application models

Abstract

This invention discloses a semantically enhanced scheduling decision-making system and method for geospatial application models. A multimodal scheduling sample generation module analyzes the scheduling indicators and scheduling relationships of geospatial application models from multi-source cluster trajectories, organizes them according to a hierarchical indicator model, generates a scheduling indicator heat map and a scheduling affinity prompt template, and constructs a sample library after pairing. A semantically enhanced scheduling reasoning module makes scheduling decisions through a cascaded scheduling representation network and a scheduling reasoning network, generating the optimal scheduling decision solution for the geospatial application model. An environmental feedback scheduling compensation module implements scheduling compensation for scheduling imbalance tasks, updates the scheduling compliance factors of relevant samples in the sample library, and simultaneously mines difficult negative samples to support efficient iteration of the scheduling representation network and the scheduling reasoning network. This invention can achieve effective management of containerized resources of large-scale geospatial application models and has strong operability and practical value.

Description

Technical Field

[0001] The present invention relates to geographic information services and cloud computing, and in particular to a semantically enhanced scheduling decision system and method for a geographic space application model. Background Art

[0002] Geospatial application models are a type of quantitative application model based on geographic information system theory and methods. By coupling multi-source, heterogeneous geospatial data with domain knowledge, they enable analysis, simulation, and deduction of geospatial phenomena, spatial interactions, and the evolutionary mechanisms of complex processes. Compared to general-purpose models in fields such as natural language processing and image classification and detection, geospatial application models are characterized by their wide range of sources, strong heterogeneity, complex spatial relationships, and spatiotemporal dynamics. They cover a wide range of applications, including feature extraction, spatial analysis, and planning assistance.

[0003] Due to the diversity and computational complexity of geospatial application models, existing scheduling systems and methods still face significant challenges in meeting model resource requirements and improving model application effectiveness. Most existing methods focus on optimizing task-side scheduling strategies, using heuristic, swarm intelligence, and machine learning scheduling methods. However, these methods lack an effective representation of the relationship between containerized resources and tasks, as well as a deep understanding of scheduling semantics, which limits the accuracy of scheduling reasoning and execution of geospatial application model containerized tasks. At the same time, existing commercial-grade scheduling frameworks (such as centralized, two-level, and state-sharing scheduling frameworks) focus more on generality and lack adaptive solutions for the characteristics and requirements of geospatial application model containerized tasks. In particular, they are difficult to effectively analyze multi-instance and multi-dependent task scheduling scenarios, which affects resource utilization efficiency and service application effectiveness.

[0004] In recent years, cross-modal contrastive learning methods based on deep networks have achieved great success in the field of computer vision. They leverage the similarities and complementarities between multimodal data to construct a shared feature space, providing new insights into the understanding of complex scheduling semantic relationships for geospatial application models. However, due to the diversity and complexity of geospatial application models, there is currently a lack of large-scale aligned image-text samples to support efficient model training, and sample sparsity limits the application and exploration of these methods in the fields of geographic information services and cloud computing. In summary, there is currently a lack of semantically enhanced scheduling decision methods and their systematic research for geospatial application models in related fields. That is, based on the establishment of fine-grained scheduling semantic representations, an integrated solution for containerized task scheduling reasoning for geospatial application models is provided. Summary of the Invention

[0005] The present invention aims to provide a semantically enhanced scheduling decision system and method for a geographic space application model.

[0006] The technical solution to realize the present invention is: a semantically enhanced scheduling decision system for geospatial application models, including a multimodal scheduling sample generation module, a semantically enhanced scheduling reasoning module, and an environmental feedback scheduling compensation module, wherein:

[0007] The multimodal scheduling sample generation module analyzes the scheduling indicators and scheduling relationships of the geospatial application model from multi-source cluster trajectories, organizes them according to a hierarchical indicator model, generates scheduling indicator heat maps and scheduling affinity prompt templates, and builds a sample library after pairing them for downstream task scheduling decisions;

[0008] The semantically enhanced scheduling reasoning module receives paired samples and makes scheduling decisions through a cascaded scheduling representation network and scheduling reasoning network to generate the optimal scheduling decision solution for the geospatial application model. The scheduling representation network is used to extract and fuse the characteristics of the multimodal geospatial application model's operating environment to generate a high-level scheduling semantic representation. The scheduling reasoning network outputs the high-level scheduling semantic representation as the optimal scheduling action for the geospatial application model in a containerized distributed cluster.

[0009] The environmental feedback scheduling compensation module is used for scheduling execution observation and scheduling compliance evaluation of containerized tasks in geospatial application models, implements scheduling compensation for scheduling imbalance tasks, updates the scheduling compliance factors of relevant samples in the sample library, and simultaneously mines difficult negative samples to support efficient iteration of the scheduling representation network and scheduling inference network.

[0010] A semantically enhanced scheduling decision method for geospatial application models is implemented by constructing the semantically enhanced scheduling decision system for geospatial application models. The specific method is as follows:

[0011] Using a multimodal scheduling sample generation module, we analyze the scheduling indicators and scheduling relationships of geospatial application models from multi-source cluster trajectories. We then organize these into a hierarchical indicator model to generate a scheduling indicator heat map and scheduling affinity prompt templates. After pairing, we construct a sample library for downstream task scheduling decisions.

[0012] The semantically enhanced scheduling reasoning module receives paired samples and makes scheduling decisions through a cascaded scheduling representation network and scheduling reasoning network to generate the optimal scheduling decision solution for the geospatial application model. The scheduling representation network is used to extract and fuse the characteristics of the multimodal geospatial application model's operating environment to generate a high-level scheduling semantic representation. The scheduling reasoning network outputs the high-level scheduling semantic representation as the optimal scheduling action for the geospatial application model in a containerized distributed cluster.

[0013] The environmental feedback scheduling compensation module is used to observe the scheduling execution and evaluate the scheduling compliance of the containerized tasks of the geospatial application model, implement scheduling compensation for scheduling imbalance tasks, update the scheduling compliance factors of relevant samples in the sample library, and simultaneously mine difficult negative samples to support the efficient iteration of the scheduling representation network and the scheduling reasoning network.

[0014] 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, the method for semantically enhanced scheduling decision-making for geospatial application models is implemented to realize semantically enhanced scheduling decision-making for geospatial application models.

[0015] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the method for semantically enhanced scheduling decision-making for geospatial application models is implemented to realize semantically enhanced scheduling decision-making for geospatial application models.

[0016] Compared with the existing technology, the present invention has the following significant advantages: a multimodal scheduling environment representation method is proposed for the characteristics and needs of geographic spatial application models, which makes up for the lack of adaptability of traditional indicator-based scheduling models in the field of geographic information. A semantically enhanced geographic application model scheduling framework is constructed, and by combining the contrastive learning scheduling representation network with the reinforcement learning scheduling reasoning network, high-order representation of containerized tasks and precise reasoning of scheduling actions are achieved, thereby improving the accuracy of geospatial application model scheduling. Based on the above effects, the present invention can achieve effective management of large-scale geospatial application model containerized resources in engineering applications in the fields of geographic information services and cloud computing, and has strong operability and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is an architectural diagram of the semantically enhanced scheduling decision system for geospatial application models of the present invention.

[0018] Figure 2 It is a flow chart of the semantically enhanced scheduling decision method for geographic space application models of the present invention.

[0019] Figure 3 This is the architecture diagram of the multimodal scheduling sample generation module.

[0020] Figure 4 This is the architecture diagram of the semantically enhanced scheduling reasoning module.

[0021] Figure 5 This is a schematic diagram of a two-stage reinforcement learning engine.

[0022] Figure 6 It is a schematic diagram of the template rule.

[0023] Figure 7 This is a schematic diagram of an affinity scheduling hint template. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0025] The structure of the semantic enhancement scheduling decision system for geographic space application model of the present invention is as follows: Figure 1 As shown, the overall process is as follows Figure 2 The semantically enhanced scheduling decision system for geospatial application models consists of a multimodal scheduling sample generation module, a semantically enhanced scheduling reasoning module, an environmental feedback scheduling compensation module, and a geospatial application model operating environment, namely a containerized distributed cluster. This system implements an integrated geospatial application model scheduling solution that integrates scheduling representation reasoning, scheduling execution observation, and scheduling feedback compensation. Specifically:

[0026] a. The multimodal scheduling sample generation module analyzes the scheduling indicators and scheduling relationships of the geospatial application model from multi-source cluster trajectories, organizes them according to the hierarchical indicator model, generates scheduling indicator heat maps and scheduling affinity prompt templates, and builds a sample library after pairing to make downstream task scheduling decisions.

[0027] b. The semantically enhanced scheduling reasoning module receives paired samples and performs scheduling representation and reasoning through a cascaded scheduling representation network and scheduling reasoning network, generating the optimal scheduling decision plan for the geospatial application model. Simultaneously, this module undergoes two-stage reinforcement learning engine optimization training: the first stage utilizes a large-scale offline sample library for generalization training, and the second stage utilizes an online sample library for scheduling compliance evaluation for fine-tuning.

[0028] c. The environmental feedback scheduling compensation module conducts observation scheduling action execution and scheduling compliance evaluation on the containerized tasks of the geospatial application model, performs scheduling compensation, i.e. dynamic scheduling, for containerized tasks with scheduling imbalance, and simultaneously mines a large number of difficult negative samples to support the efficient iteration of the scheduling representation / inference network.

[0029] The following combination Figure 2-5 The composition and function of each module are described in detail.

[0030] (1) Multimodal Scheduling Sample Generation

[0031] The multimodal scheduling sample generation module includes four submodules: multi-source cluster trajectory access, hierarchical indicator model organization, scheduling indicator heat map generation, and scheduling affinity prompt template generation. It is used to generate paired image and text samples from multi-source cluster trajectories, and build a large-scale offline sample library and online sample library to support downstream containerized task scheduling decisions. In particular, the scheduling indicator heat map is composed of a resource indicator heat map and a task indicator heat map, which is used to represent the containerized distributed scheduling environment. The scheduling affinity prompt template is composed of four types of template rules, which is used to represent the containerized task scheduling relationship of the geospatial application model. Figure 1 、 3 The specific implementation steps are as follows:

[0032] (1) Multi-source cluster trajectory access

[0033] These cluster trajectories are both offline, open-source, and real-time local cluster trajectories for large-scale geospatial application model containerization tasks. The open-source cluster trajectories are used to build an offline sample library to support generalization training of the downstream scheduling representation network; the local cluster trajectories are used to build an online sample library to support downstream scheduling inference network decision-making and fine-tuning of the scheduling representation network.

[0034] Chain storage consistency is used to access two types of cluster trajectory data, supporting the playback of the scheduling process at any scheduling time and the mining of scheduling value information. Chain storage is formally expressed as:

[0035]

[0036] in, Respectively represent scheduling The cluster state at a certain moment, the mapping between scheduled containerized tasks and computing resource nodes, and the cluster state at the next moment. The cluster state data includes configuration information and usage information of computing resource nodes, containerized tasks, and container instances.

[0037] (2) Hierarchical indicator model organization

[0038] The hierarchical indicator model is used to structure the scheduling indicators and scheduling relationships of the geospatial application model, and includes four levels from top to bottom: cluster indicators, computing resource node indicators, task indicators, and container indicators.

[0039] Cluster indicators are scheduling indicators related to containerized distributed clusters, including the number of computing resource nodes in the cluster. , the number of cluster computing resource types , the number of cluster containerized task types , the number of cluster containerized tasks , cluster computing resource quota , and cluster computing resource utilization ,in Indicates the computing resource type number;

[0040] Resource node indicators include scheduling indicators related to each computing resource node in the containerized distributed cluster, and resource quotas of computing resource nodes. ,in Indicates the computing resource node number; the resource utilization of the computing resource node , the number of containerized tasks for computing resource nodes , containerized task type of computing resource node , and the task occupancy rate of computing resource nodes ,in Indicates the containerized task type number;

[0041] Task indicators are scheduling indicators corresponding to the containerized tasks of the geospatial application model, including resource request volume , the number of container instances , task dependencies ,in Indicates the containerized task number;

[0042] Container metrics include container resource requests , and container scheduling relationships ,in Indicates the container instance ID.

[0043] (3) Construction of scheduling indicator heat map samples

[0044] The scheduling indicator heat map is used for the graphical representation of the operation and scheduling environment status of the geospatial application model. It is composed of a heat map of resource indicators that are continuous in time and space. and task indicator heat map sequence Composition. Time continuous and Indicates the state of the scheduling environment, expressed as:

[0045]

[0046] in, For scheduling The scheduling environment state observation at the moment is the scheduling indicator heat map, which is represented as a sequence of resource indicator heat maps and task indicator heat map Sequence combination; Represent the length, width and number of channels of the heat map respectively. , where the horizontal axis represents the computing resource node number , the vertical axis represents the computing resource type , pixel value Represents a computing resource node Computing resource type Utilization rate . Similarly, for any , where the horizontal axis represents the computing resource node number , the vertical axis represents the containerized task type , pixel value Indicates the occupancy rate of the corresponding type of containerized tasks on the computing resource node .

[0047] Furthermore, the sequence color map is used to Mapped to three-channel pixel values ,according to The value ranges from 0 to 1, and the color mapping goes from light to dark, indicating a linear change in resource utilization. Expressed as:

[0048]

[0049] in, The function is used to convert a floating point number into the corresponding integer value.

[0050] Use sequence color mapping to Mapped to three-channel pixel values ,according to The value ranges from 0 to 1, and the color mapping is from light to dark, indicating a linear change in task occupancy. Task indicator heat map pixel value Expressed as:

[0051]

[0052] (4) Construction of scheduling affinity prompt template

[0053] The affinity prompt template converts the affinity scheduling semantic relationship in the indicator file into a form that is easy for the scheduling model to understand through four types of template rules. The affinity scheduling semantic relationship includes resource affinity scheduling semantic relationship and task affinity scheduling semantic relationship.

[0054] Design template rules R-1, R-2 to extract resource affinity scheduling semantic relationship, T-1, T-2 to extract task affinity scheduling semantic relationship, the template rule definition is as follows Figure 6 As shown, R-1 extracts the basic demand for various computing resources for the containerized task of the geospatial application model, including: CPU request , memory request amount , disk request volume and bandwidth requests ; R-2 extracts the maximum demand for various computing resources for the containerized task of the geospatial application model, including: CPU limit , memory limit , disk quota and bandwidth limits ; T-1 extracts the dependency relationship of the containerized tasks of the geospatial application model and filters out the identifiers of all dependent containerized tasks; T-2 extracts the multi-instance deployment relationship of the containerized tasks of the geospatial application model and filters out the task identifiers of all container instances.

[0055] Furthermore, using 7 tokens ( ) to identify resource affinity scheduling semantic relationships and task affinity scheduling semantic relationships, and generate affinity scheduling prompt templates ,like Figure 7 As shown, It is the unique identifier of the containerized task. is the resource affinity attribute, This is the affinity attribute of the containerized task.

[0056] 2. Semantic Enhanced Scheduling Reasoning

[0057] The semantically enhanced scheduling reasoning module includes a scheduling representation network, a scheduling reasoning network, and a two-stage reinforcement learning engine, which is used to receive upstream geospatial application model samples, and make scheduling decisions through the cascade scheduling representation network and scheduling reasoning network to generate the optimal geospatial application model containerized task scheduling decision plan. The scheduling representation network is composed of an asymmetric visual encoder, a text encoder, and a multimodal encoder, which is used for multimodal geospatial application model scheduling environment feature extraction and fusion to generate high-order scheduling semantic representation; the scheduling reasoning network is a deep reinforcement learning network, which maps the high-order scheduling semantic representation to the optimal scheduling action of the geospatial application model in the containerized distributed cluster. The scheduling representation / reasoning network is optimized and trained by a two-stage reinforcement learning engine. The first stage uses large-scale offline sample generalization training, and the second stage uses online samples that have undergone scheduling compliance evaluation for fine-tuning. Reference Figure 1 、 4 5. The specific implementation steps are as follows:

[0058] (1) Comparative Learning Scheduling Representation

[0059] The scheduling representation network consists of an asymmetric visual encoder, a text encoder, and a multimodal encoder. The visual encoder extracts geospatial application model scheduling environment features from the scheduling indicator heat map, the text encoder extracts scheduling affinity features from the scheduling context, and the multimodal encoder fuses the view encoder and text encoder to output high-level semantic features for scheduling decisions in the downstream scheduling inference network.

[0060] Specifically, the visual encoder uses a lightweight ResNet-18 as the backbone network, removes the last classifier layer, and then adds a global average pooling layer to extract global features with a dimension of 512. The fully connected layer maps the global features to 128 dimensions. Finally, a batch normalization layer is added to improve the network convergence and generalization capabilities.

[0061] The text encoder first uses template rules to extract resource affinity and task affinity attributes from various indicator data sources to fill in the affinity scheduling prompt template. Then, the prompt template is converted into a 512-dimensional vector representation using a pre-trained Doc2Vec model. Finally, this vector is fed into an MLP network with two hidden layers to produce a 128-dimensional affinity feature output.

[0062] The multimodal encoder uses a linear projection layer to merge the output vector representations of the visual encoder and text encoder into the multimodal space, generating 256-dimensional semantic features that are output to the scheduling reasoning network for scheduling decisions.

[0063] (2) Reinforcement Learning Scheduling Reasoning

[0064] The scheduling inference network has a structure with 3 hidden layers of exactly the same depth Network and Target Network composition, depth The network uses the 256-dimensional semantic feature vector output by the multimodal encoder as state input and outputs the scheduling probability distribution of each computing resource node in the current state; the goal Network periodic synchronization depth The network parameters output the optimal scheduling action, that is, the computing resource node for the optimal scheduling of the current geospatial application model containerized task. The state space, action space, and reward function of the inference network are specifically designed as follows:

[0065] 1) State Space

[0066] The state space is the collection of states of the scheduling reasoning network in the geospatial application model operating environment, i.e., the containerized distributed cluster. Assume that the scheduling representation network is represented as , then the state space of the scheduling reasoning network is expressed as:

[0067]

[0068] in, They represent the scheduling indicator heat map and scheduling affinity prompt template respectively.

[0069] 2) Action Space

[0070] The action space is the set of scheduling actions output by the scheduling reasoning network under a given state. Assuming that the containerized distributed cluster consists of Computing resource nodes, the currently scheduled containerized tasks have container instances, each of which may be scheduled to any computing resource node in the distributed system. The action space of the scheduling reasoning network is expressed as:

[0071]

[0072] Among them, the matrix elements , 1 represents a container instance Scheduling to computing resource nodes , 0 indicates a container instance Not scheduled to computing resource nodes .

[0073] 3) Reward Function

[0074] The reward function is used to evaluate the quality of scheduling actions, thereby guiding the scheduling reasoning network to optimize scheduling in the direction of obtaining the maximum cumulative reward. The scheduling decision system of the present invention aims to optimize the efficiency of resource balanced utilization. The reward function of the scheduling reasoning network is expressed as:

[0075]

[0076] in, Indicates the number of cluster computing resource nodes, Indicates the number of cluster computing resource types; Represents a computing resource node The average resource utilization rate, Represents a computing resource node Resource utilization balance, coordination coefficient Used to balance the two;

[0077] (3) Two-stage scheduling enhancement engine

[0078] The two-stage scheduling enhancement engine is placed between the sample library and the cascaded scheduling representation network and scheduling reasoning network. In the first stage, the scheduling representation network is trained using large-scale offline samples. In the second stage, the scheduling representation network is fine-tuned using online samples that have undergone scheduling compliance evaluation and the scheduling reasoning network is trained. The specific implementation steps are as follows:

[0079] Phase 1: Scheduling Representation Network Generalization Training

[0080] 1) Random sampling from a large offline sample library Mini-batch sample pairs ,in is the image sample, i.e. the scheduling index heat map, For each of the text, that is, the scheduling affinity template. Enhanced image samples are obtained by enhancement ,form Sample pairs Each image sample With enhanced image samples Form a positive sample pair, and the remaining image samples form a negative sample pair. Each image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs, and each enhanced image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs.

[0081] 2) Using image contrast loss Train the visual encoder to extract visual features from the scheduling indicator heat map, using image-text contrast loss Jointly train the visual encoder and text encoder to align visual features with affinity text features, using image-text matching loss The joint representation of scheduling semantic features is learned in the multimodal space, and the relevant definitions are as follows:

[0082]

[0083] in, , , Represents image samples respectively , enhanced image samples , and negative samples The visual feature vector of is the indicator function, is the cosine similarity, is an exponential function, is a logarithmic function, is the temperature coefficient, is the number of mini-batch samples.

[0084]

[0085] in, Represents image samples respectively Visual feature vector and affinity text The meanings of other symbols are the same as those in formula (6).

[0086]

[0087] in, Represents an image sample and affinity text Matching true labels, Representing visual features Probability of matching with text features.

[0088] Phase 2: Fine-tuning the scheduling representation network and training the scheduling inference network

[0089] 1) Fine-tuning the Scheduling Representation Network

[0090] From the online and offline sample libraries, random sampling is performed each time Mini-batch sample pairs ,in is the image sample, i.e. the scheduling index heat map, For each of the text, that is, the scheduling affinity template. Enhanced image samples are obtained by enhancement ,form Sample pairs Each image sample With enhanced image samples Form a positive sample pair, and the remaining image samples form a negative sample pair. Each image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs, and each enhanced image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs.

[0091] Image contrast loss, image-text contrast loss, and image-text matching loss are used to fine-tune the network parameters of the view encoder, text encoder, and multimodal encoder until the error converges.

[0092] 2) Scheduling Inference Network Training

[0093] Each iteration randomly samples from the online sample library Mini-batch trajectories ,in Represents scheduling The cluster status at the moment, Represents the mapping relationship between scheduled containerized tasks and computing resource nodes, represents the scheduling reward, Represents scheduling The cluster status at the moment. Calculate the expected square of the time series difference error , update the depth using stochastic gradient descent Network parameters. Defined as:

[0094]

[0095] , (11);

[0096] in, Indicates depth Network parameters, represents the timing difference error, represents the reward function (see Formula 7a for details), Indicates the target network, express The network parameters, Indicates the cluster status and scheduling action at the next moment; Indicates seeking Maximizing scheduling actions , is the discount rate.

[0097] Every iterations, synchronize once Network parameters To the target network parameter , used to reduce The fluctuation of the objective function in the network enhances the learning robustness. It is a network hyperparameter and is set to an integer value greater than 1 according to the specific scheduling scenario.

[0098] The environmental feedback scheduling compensation module consists of a scheduling compliance evaluation module and a difficult negative sample mining module, which are used for scheduling evaluation feedback, scheduling compensation and scheduling network optimization iteration of geospatial application model containerized tasks. The scheduling compliance evaluation module, by establishing a timed environmental scanning mechanism, periodically observes scheduling execution and performs compliance evaluation, screens geospatial application model containerized tasks with scheduling imbalance in real time and performs dynamic scheduling compensation, establishes a special event response mechanism, quickly handles cluster emergencies such as computing resource node downtime and service crashes, and promptly migrates containerized tasks on abnormal computing resource nodes to other valid computing resource nodes to ensure service stability and reliability. The difficult negative sample mining module evaluates the difficulty of identification between negative samples through multiple criteria, and mines a large number of difficult negative samples to support efficient iteration of scheduling representation / inference networks. The specific implementation steps are as follows:

[0099] (1) Scheduling compliance evaluation

[0100] Divide the geospatial application model containerization tasks into long-running services based on the running cycle , batch jobs , periodic operations and workflow .for , using the average service level target violation rate Evaluate scheduling performance; for 、 and , using the average job completion time Evaluate scheduling performance. The relevant definitions are as follows:

[0101]

[0102] in, They represent the service level objective violation rate threshold and job completion time threshold respectively. For containerized tasks Type, and Respectively response time and response time threshold, for gather, for 、 and gather, Represents a set and capacity.

[0103] The scheduling compliance evaluation module has two built-in scheduling evaluation mechanisms: scheduled environment scanning and special event response. Analyze the scheduling environment periodically at time intervals, use formulas (12a) and (12c) to observe the scheduling execution and conduct compliance evaluation, and Indicator exceeds threshold The containerized tasks are pushed into the rescheduling queue for dynamic scheduling compensation; the special event response mechanism observes cluster emergencies such as computing resource node downtime and service crash in real time, and pushes all containerized tasks of abnormal computing resource nodes into the rescheduling queue.

[0104] (2) Difficult Negative Sample Mining

[0105] Difficult negative samples refer to negative samples that are difficult to be correctly identified by the scheduling representation network. These samples have low recognition and high similarity, which limits the ability of scheduling semantic representation.

[0106] Specifically, for any image sample All negative sample sets , the difficulty of negative samples for image samples is evaluated by approximating the ideal solution ranking method TOPSIS:

[0107]

[0108] in, Represents negative samples With image samples The distance between them, that is, the negative samples With image samples The absolute value of the difference in the number of containerized tasks scheduled between them; and Represent negative samples respectively The average resource utilization and resource imbalance of the cluster at the sampling time. The smaller, The bigger, The smaller the negative sample difficulty The bigger it is, the smaller it is.

[0109] Furthermore, the penalty factor Image contrast loss Scaling processing amplifies the proportion of difficult negative sample loss in the adjustment of scheduling representation network parameters, guiding the network to learn in the direction of correctly identifying difficult negative samples to improve the accuracy of scheduling semantic representation.

[0110]

[0111] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the semantically enhanced scheduling decision method for geospatial application models is implemented to realize semantically enhanced scheduling decisions for geospatial application models.

[0112] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the method for semantically enhanced scheduling decision-making for geospatial application models is implemented to realize semantically enhanced scheduling decision-making for geospatial application models.

[0113] In summary, the semantically enhanced scheduling decision system for geospatial application models realizes an integrated solution of affinity semantically enhanced scheduling representation reasoning, scheduling execution observation, and scheduling feedback compensation in containerized distributed clusters through multimodal scheduling sample generation, semantically enhanced scheduling reasoning, and environmental feedback scheduling compensation. It makes up for the lack of adaptability of traditional indicator-based scheduling models in the field of geographic information and improves the accuracy of containerized task scheduling of geospatial application models.

[0114] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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.

[0115] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A semantically enhanced scheduling decision system for geospatial application models, characterized by: It includes a multimodal scheduling sample generation module, a semantically enhanced scheduling reasoning module, and an environmental feedback scheduling compensation module, among which: The multimodal scheduling sample generation module analyzes the scheduling indicators and scheduling relationships of the geospatial application model from multi-source cluster trajectories, organizes them according to a hierarchical indicator model, generates scheduling indicator heat maps and scheduling affinity prompt templates, and builds a sample library after pairing them for downstream task scheduling decisions; The semantically enhanced scheduling reasoning module receives paired samples and makes scheduling decisions through a cascaded scheduling representation network and scheduling reasoning network to generate the optimal scheduling decision solution for the geospatial application model. The scheduling representation network is used to extract and fuse the characteristics of the multimodal geospatial application model's operating environment to generate a high-level scheduling semantic representation. The scheduling reasoning network outputs the high-level scheduling semantic representation as the optimal scheduling action for the geospatial application model in a containerized distributed cluster. The environmental feedback scheduling compensation module is used for scheduling execution observation and scheduling compliance evaluation of containerized tasks in geospatial application models. It implements scheduling compensation for scheduling imbalance tasks, updates the scheduling compliance factors of relevant samples in the sample library, and mines difficult negative samples to support the iteration of the scheduling representation network and scheduling inference network. in, Multi-source cluster trajectories include open-source cluster trajectories and local cluster trajectories. The open-source cluster trajectories are used to build an offline sample library to support the generalization training of the downstream scheduling representation network; the local cluster trajectories are used to build an online sample library to support the downstream scheduling reasoning network decision-making and scheduling representation network fine-tuning. Scheduling indicators include cluster indicators, resource node indicators, task indicators, and container indicators. Cluster indicators are scheduling indicators related to containerized distributed clusters, including the number of cluster computing resource nodes. , the number of cluster computing resource types , the number of cluster containerized task types , the number of cluster containerized tasks , cluster computing resource quota , and cluster computing resource utilization ,in Indicates the computing resource type number; resource node indicators include scheduling indicators related to each computing resource node in the containerized distributed cluster, including computing resource node quotas ,in Indicates the computing resource node number; the resource utilization of the computing resource node , the number of containerized tasks for computing resource nodes , computing resource node containerized task type , and computing resource node task occupancy ,in Indicates the containerized task type number; the task indicator is the scheduling indicator corresponding to the geospatial application model containerized task, including the resource request amount , the number of container instances , task dependencies ,in Indicates the containerized task number; container metrics include container resource requests , and container scheduling relationships ,in Indicates the container instance ID.

2. The semantically enhanced scheduling decision system for geospatial application models according to claim 1 is characterized in that: The multimodal scheduling sample generation module includes a multi-source cluster trajectory access submodule, a hierarchical indicator model organization submodule, a scheduling indicator heat map generation submodule, and a scheduling affinity prompt template generation submodule, among which: The multi-source cluster trajectory access submodule is used to access open source cluster trajectories and local cluster trajectories of large-scale geospatial application model containerization tasks. The open source cluster trajectories are used to build an offline sample library to support the generalization training of the downstream scheduling representation network; the local cluster trajectories are used to build an online sample library to support the downstream scheduling reasoning network decision-making and scheduling representation network fine-tuning. The hierarchical indicator model organization submodule is used to structure the scheduling indicators and scheduling relationships of the geospatial application model. From top to bottom, it includes four levels: cluster indicators, resource node indicators, task indicators, and container indicators. The scheduling indicator heat map generation submodule is used for the graphical representation of the operation of the geospatial application model and the scheduling environment status, which is composed of the time and space continuous resource indicator heat map and task indicator heat map sequence composition; The scheduling affinity prompt template generation sub-module converts the affinity scheduling semantic relationship in the indicator file into a form that is easy to understand for the scheduling model through four types of template rules. The affinity scheduling semantic relationship includes resource affinity scheduling semantic relationship and task affinity scheduling semantic relationship. The resource affinity scheduling semantic relationship is extracted through template rules R-1 and R-2, and the task affinity scheduling semantic relationship is extracted through template rules T-1 and T-2. Among them, R-1 extracts the basic demand of the geospatial application model containerization task for various computing resources, including: CPU request amount, memory request amount, disk request amount and bandwidth request amount; R-2 extracts the maximum demand of the geospatial application model containerization task for various computing resources, including: CPU limit, memory limit, disk limit and bandwidth limit; T-1 extracts the dependency relationship of the geospatial application model containerization task and filters out the task identifiers of all dependent tasks; T-2 extracts the multi-instance deployment relationship of the geospatial application model containerization task and filters out the task identifiers of all container instances.

3. The semantically enhanced scheduling decision system for geospatial application models according to claim 2 is characterized in that: use A heat map of resource indicators with continuous time and indicator heat map sequence Indicates the state of the scheduling environment, expressed as: ； in, Respectively represent scheduling Observation of the scheduling environment status at each moment, time series of resource indicator heat maps, and time series of task indicator heat maps; Represent the length, width and number of channels of the heat map respectively; for any , where the horizontal axis represents the computing resource node number , the vertical axis represents the computing resource type , pixel value Represents a computing resource node Computing resource type Utilization rate ; For any , where the horizontal axis represents the computing resource node number , the vertical axis represents the containerized task type , pixel value Represents a computing resource node The above type is Containerized task share ; Use sequence color mapping to Mapped to three-channel pixel values ,according to The value ranges from 0 to 1, and the color mapping is from light to dark, indicating a linear change in resource utilization. The pixel value of the resource indicator heat map Expressed as: ； in, The function is used to convert floating point numbers into corresponding integer values; Use sequence color mapping to Mapped to three-channel pixel values ,according to The value ranges from 0 to 1, and the color mapping is from light to dark, indicating the linear change of containerized task occupancy rate. The pixel value of the task indicator heat map Expressed as: 。 4. The semantically enhanced scheduling decision system for geospatial application models according to claim 2, characterized in that: The semantically enhanced scheduling reasoning module includes a contrastive learning scheduling representation network and a reinforcement learning scheduling reasoning network, wherein: The scheduling representation network consists of an asymmetric visual encoder, a text encoder, and a multimodal encoder. The visual encoder extracts the scheduling environment features of the geospatial application model from the scheduling indicator heat map. The text encoder extracts scheduling affinity features from the scheduling context. The multimodal encoder fuses the view encoder and text encoder to output high-order semantic features for scheduling decisions in the downstream scheduling reasoning network. The visual encoder uses ResNet-18 as the backbone network, removing the last classifier layer. A global average pooling layer, a fully connected layer, and a batch normalization layer are then added in sequence. The global average pooling layer extracts global features with a dimension of 512. The fully connected layer maps the global features to 128 dimensions. Finally, a batch normalization layer is added to improve network convergence and generalization capabilities. The text encoder first extracts resource affinity and task affinity attributes through template rules to fill in the affinity hint template. The affinity hint template is converted into a 512-dimensional vector representation using the pre-trained Doc2Vec model and fed into an MLP network with two hidden layers to generate a 128-dimensional affinity feature output. The multimodal encoder uses a linear projection layer to merge the output vector representations of the visual encoder and text encoder into the multimodal space, generating 256-dimensional semantic features that are output to the scheduling reasoning network for scheduling decisions; The scheduling inference network has a structure with 3 hidden layers of exactly the same depth Network and Target Network composition, depth The network uses the 256-dimensional semantic feature vector output by the multimodal encoder as state input and outputs the scheduling probability distribution of each computing resource node in the current state; the goal is Network periodic synchronization depth The network parameters output the optimal scheduling action, that is, the computing resource node for the optimal scheduling of the containerized task of the current geospatial application model. The specific designs of the state space, action space, and reward function are as follows: 1) State Space The state space is the set of all state inputs of the scheduling reasoning network in the geospatial application model operating environment, i.e., the containerized distributed cluster. Assume that the scheduling representation network is represented as , then the state space of the scheduling reasoning network is expressed as: ； in, They represent the scheduling environment observation and scheduling affinity prompt template respectively; 2) Action Space The action space is the set of scheduling actions output by the scheduling reasoning network in a given state. The scheduling action is the mapping relationship between containerized tasks and resource nodes. Assume that the containerized distributed cluster consists of Computing resource nodes, the currently scheduled containerized tasks have container instances, each of which may be scheduled to any computing resource node in the distributed system. The action space of the scheduling reasoning network is expressed as: ； Among them, the matrix elements , , 1 represents the container instance of the containerized task Scheduling to computing resource nodes , 0 indicates a container instance Not scheduled to computing resource nodes ; 3) Reward Function The reward function is used to evaluate the quality of the scheduling action, thereby guiding the scheduling reasoning network to optimize the scheduling in the direction of obtaining the maximum cumulative reward. The reward function of the scheduling reasoning network is expressed as: ； in, Indicates the number of computing resource nodes, Indicates the number of computing resource types on the computing resource node; Represents a computing resource node The average resource utilization rate, Represents a computing resource node Resource utilization balance, coordination coefficient Used to balance the two.

5. The semantically enhanced scheduling decision system for geospatial application models according to claim 1 is characterized in that: The semantically enhanced scheduling reasoning module also includes a two-stage reinforcement learning engine. In the first stage, the scheduling representation network is trained by generalizing offline samples. In the second stage, the scheduling representation network is fine-tuned using online samples that have undergone scheduling conformity evaluation and the scheduling reasoning network is trained. The specific implementation steps are as follows: Phase 1: Scheduling Representation Network Generalization Training 1) Random sampling from the offline sample library Mini-batch sample pairs ,in is the image sample, i.e. the scheduling index heat map, For the text sample, the affinity template is scheduled, and for each image sample Enhanced image samples are obtained by enhancement ,form Sample pairs ; Each image sample With enhanced image samples Form a positive sample pair, and the remaining image samples form a negative sample pair; each image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs, and each enhanced image sample Paired with text samples Form a positive sample pair, and the remaining Text samples form negative sample pairs; 2) Using image contrast loss Train the visual encoder to extract visual features from the scheduling indicator heat map, using image-text contrast loss Jointly train the visual encoder and text encoder to align visual features with affinity text features, using image-text matching loss The joint representation of scheduling semantic features is learned in the multimodal space, and the relevant definitions are as follows: ； in, , , Represents image samples respectively , enhanced image samples , and negative samples The visual feature vector of is the indicator function, is the cosine similarity, is an exponential function, is a logarithmic function, is the temperature coefficient, is the number of mini-batch samples; ； in, Represents image samples respectively Visual feature vectors and text samples The text feature vector of ； in, Represents an image sample With text sample Matching true labels, Representing visual features Probability of matching with text features; Phase 2: Fine-tuning the scheduling representation network and training the scheduling inference network 1) Fine-tuning the Scheduling Representation Network From the online and offline sample libraries, random sampling is performed each time Mini-batch sample pairs ,in is the image sample, i.e. the scheduling index heat map, For the text sample, the affinity template is scheduled, and for each image sample Enhanced image samples are obtained by enhancement ,form Sample pairs , each image sample With enhanced image samples Form a positive sample pair, and the remaining image samples form a negative sample pair; each image sample Paired with text samples Form a positive sample pair, and the remaining The text samples form negative sample pairs, and each enhanced image sample Paired with text samples Form a positive sample pair, and the remaining Text samples form negative sample pairs; Using image contrast loss, image-text contrast loss, and image-text matching loss, fine-tune the network parameters of the view encoder, text encoder, and multimodal encoder until the error converges; 2) Scheduling Inference Network Training Each iteration randomly samples from the online sample library Mini-batch trajectories ,in Represents scheduling The cluster status at the moment, Represents the mapping relationship between scheduled containerized tasks and computing resource nodes, represents the scheduling reward, Represents scheduling The cluster status at the moment, calculate the expectation of the square of the time series difference error , update the depth using stochastic gradient descent Network parameters; Defined as: ； in, Indicates depth Network parameters, represents the timing difference error, represents the reward function, Indicates the target network, express The network parameters, Indicates the cluster status and scheduling action at the next moment; Indicates seeking Maximizing scheduling actions , is the discount rate; Every iterations, synchronize once Network parameters To the target network parameter , used to reduce Fluctuations in the objective function of the network, enhancing learning robustness, It is a network hyperparameter and is set to an integer value greater than 1 according to the specific scheduling scenario.

6. The semantically enhanced scheduling decision system for geospatial application models according to claim 1 is characterized in that: The environmental feedback scheduling compensation module includes a scheduling compliance evaluation module. By establishing a timed environmental scanning mechanism, it periodically observes scheduling execution and conducts compliance evaluation. It screens out unbalanced geospatial application model containerized tasks in real time and performs dynamic scheduling compensation. It also establishes a special event response mechanism to quickly handle computing resource node downtime and service crashes, and promptly migrates containerized tasks on abnormal computing resource nodes to other valid computing resource nodes to ensure service stability and reliability. Divide the geospatial application model containerization tasks into long-running services based on the running cycle , batch jobs , periodic operations and workflow , for long-running services , using the average service level target violation rate Evaluate scheduling performance; for batch jobs , periodic operations and workflow , using the average job completion time To evaluate scheduling performance, the following definitions are provided: ； in, They represent the service level objective violation rate threshold and job completion time threshold respectively. For containerized tasks Type, and Respectively response time and response time threshold, for gather, for 、 and gather, Represents a set and capacity; Built-in two scheduling evaluation mechanisms: timed environment scanning and special event response. Timed environment scanning is used to Analyze the scheduling environment periodically at time intervals, use formulas (12a) and (12c) to observe the scheduling execution and conduct compliance evaluation, and Indicator exceeds threshold The containerized tasks are pushed into the rescheduling queue for dynamic scheduling compensation; the special event response mechanism observes the downtime of computing resource nodes and service crashes in real time, and pushes all containerized tasks on abnormal computing resource nodes into the rescheduling queue.

7. The semantically enhanced scheduling decision system for geospatial application models according to claim 6 is characterized in that: The environmental feedback scheduling compensation module includes a difficult negative sample mining module, which uses multiple criteria to evaluate the difficulty of identifying negative samples and mine difficult negative samples to support the iteration of the scheduling representation network and the scheduling reasoning network. For any image sample All negative sample sets , the difficulty of negative samples for image samples is evaluated by approximating the ideal solution ranking method TOPSIS: ； in, Represents negative samples With image samples The distance between them, that is, the negative samples With image samples The absolute value of the difference in the number of containerized tasks between them; and Represent negative samples respectively The average resource utilization and resource imbalance of the cluster at the sampling time, The smaller, The bigger, The smaller the negative sample difficulty The bigger it is, the smaller it is; Using penalty factor Image contrast loss Scaling processing: amplifies the proportion of difficult negative sample loss in the adjustment of scheduling representation network parameters, guiding the network to learn to correctly identify difficult negative samples to improve the accuracy of scheduling semantic representation; 。 8. A semantically enhanced scheduling decision method for geospatial application models, characterized in that: By constructing the semantically enhanced scheduling decision system for geospatial application models as described in any one of claims 1 to 7, semantically enhanced scheduling decision for geospatial application models is realized, and the specific method is as follows: Using a multimodal scheduling sample generation module, we analyze the scheduling indicators and scheduling relationships of geospatial application models from multi-source cluster trajectories. We then organize these into a hierarchical indicator model to generate a scheduling indicator heat map and scheduling affinity prompt templates. After pairing, we construct a sample library for downstream task scheduling decisions. The semantically enhanced scheduling reasoning module receives paired samples and makes scheduling decisions through a cascaded scheduling representation network and scheduling reasoning network to generate the optimal scheduling decision solution for the geospatial application model. The scheduling representation network is used to extract and fuse the characteristics of the multimodal geospatial application model's operating environment to generate a high-level scheduling semantic representation. The scheduling reasoning network outputs the high-level scheduling semantic representation as the optimal scheduling action for the geospatial application model in a containerized distributed cluster. The environmental feedback scheduling compensation module is used to observe the scheduling execution and evaluate the scheduling compliance of the containerized tasks of the geospatial application model, implement scheduling compensation for scheduling imbalance tasks, update the scheduling compliance factors of relevant samples in the sample library, and simultaneously mine difficult negative samples to support the iteration of the scheduling representation network and the scheduling reasoning network.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for semantically enhanced scheduling decision-making for geospatial application models according to claim 8 is implemented to realize semantically enhanced scheduling decision-making for geospatial application models.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for semantically enhanced scheduling decision-making for geospatial application models according to claim 8 is implemented to realize semantically enhanced scheduling decision-making for geospatial application models.

Images