Container orchestration-based dynamic elastic cloud practical training environment rapid construction system

Abstract

This container orchestration-based dynamic elastic cloud training environment rapid construction system comprises a cloud platform and seven related modules. The surveying and mapping training scenario template management module uses a four-layer decoupled architecture to build personalized cloud training templates; the real-time cloud training environment creation module batches single training environments or clusters; the intelligent scheduling module matches optimal computing nodes; the template call estimation module triggers pre-startup based on confidence levels; the pre-startup execution module pre-creates lightweight environments and enables idle resource reclamation or rapid activation; and the elastic scaling rules module executes elastic rule generation instructions according to priority. Tenants obtain independent, isolated resources after identity authentication. Environment construction improves efficiency through multi-task parallelism, adapting to all surveying and mapping training scenarios, achieving efficient construction and flexible adaptation of surveying and mapping training environments. Simultaneously, the forward-looking pre-startup of training environments significantly optimizes the tenant's training experience.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, specifically a rapid construction system for dynamic elastic cloud training environments based on container orchestration. Background Technology

[0002] The geographic surveying and mapping industry is a technology-intensive industry, and its training and teaching are highly specialized and unique: First, there are many types of training software with complex dependencies, covering GIS basic software, remote sensing image processing software, UAV aerial survey 3D modeling software, geodetic adjustment software, geographic big data processing platforms, etc. Different software have vastly different requirements for underlying dependency libraries, operating system versions, and computing resources, making environment configuration difficult; Second, the training data volume is huge, with a single set of aerial survey imagery or laser point cloud datasets reaching tens or even hundreds of gigabytes, and data distribution and copying are extremely time-consuming; Third, the computing power requirements vary significantly, with basic GIS operations requiring only ordinary CPU resources, while 3D modeling, remote sensing AI interpretation, and point cloud processing require high-performance GPUs and large memory support; Fourth, the training scenarios are highly cyclical, with concentrated high-load use during class time and complete idleness after class and on holidays; Fifth, a large amount of surveying and mapping training data involves classified content, requiring extremely high standards for data security and access control.

[0003] Current geographic surveying training environments mainly adopt the traditional local computer room model, which has the following core pain points: 1. Extremely high environment configuration and maintenance costs: It takes several days to configure a single class's training environment. Software version conflicts and dependency compatibility issues frequently occur between different training projects. Teachers need to spend a lot of energy solving environment problems and cannot focus on teaching content. 2. Extremely low resource utilization and high hardware investment costs: The local computer room uses fixed hardware configurations, which cannot adapt to the differentiated computing power requirements of different training projects. High-performance computing equipment is idle during off-peak hours and insufficient during peak hours. At the same time, high-specification hardware needs to be configured for each terminal, resulting in extremely high initial investment costs. TB-level training data needs to be copied repeatedly for each student, which is extremely time-consuming. In addition, the local storage and distribution of confidential data is extremely prone to data leakage, and there is a lack of effective end-to-end security control measures. 3. Extremely poor elastic scalability: When faced with sudden scenarios such as skills competitions, centralized training, and simultaneous training of multiple classes, local data centers cannot quickly expand their capacity and cannot meet the training needs of instantaneous high concurrency; current mainstream cloud training systems are mostly geared towards computer programming and general office scenarios, and have not been customized and optimized for the field of geographic surveying and mapping. They lack pre-integrated surveying and mapping-specific software images, have no optimization for the storage and sharing of surveying and mapping big data, and lack elastic scheduling strategies adapted to surveying and mapping training scenarios, thus failing to meet the professional needs of geographic surveying and mapping training.

[0004] Container orchestration technologies (such as Kubernetes) have been widely used in the cloud computing field. Their characteristics of environment consistency, rapid deployment, elastic scaling, and resource pooling management are highly compatible with the core requirements of geographic surveying and mapping training environments. However, there is currently no rapid construction system for dynamic elastic cloud training environments that is deeply customized for geographic surveying and mapping training scenarios and covers the entire process. Filling this technological gap has important industry value and application significance. Summary of the Invention

[0005] To address the aforementioned technical issues, the present invention aims to provide a rapid construction system for a dynamic and elastic cloud training environment based on container orchestration, including a cloud platform. The cloud platform is connected to a mapping training scenario template management module, a tenant information database, a real-time cloud training environment creation module, an intelligent scheduling module, a template call estimation module, a pre-startup execution module, and an elastic scaling rule module. The template management module for surveying and mapping training scenarios is used to build and store personalized cloud training templates based on customer requirements; The real-time cloud training environment creation module is used to create single training environments or training environment clusters for tenants in batches based on the personalized cloud training templates to which the tenant belongs. The intelligent scheduling module is used to perform computing node resource matching and obtain the target computing nodes for a single training environment / training environment cluster during the creation process. The template call prediction module is used to build a template call prediction model, obtain the predicted call timestamp and confidence level of the personalized cloud training template, and trigger the pre-launch of the personalized cloud training template based on the predicted call timestamp and confidence level of the personalized cloud training template. The pre-startup execution module is used to pre-create lightweight single training environments / training environment clusters based on pre-startup personalized cloud training templates, and to perform idle recycling or rapid activation operations on lightweight single training environments / training environment clusters. The elastic scaling rules module is used to execute the elastic scaling rules of the training configuration layer in the personalized cloud training template and generate elastic operation instructions.

[0006] Furthermore, the process of building personalized cloud training templates includes: Personalized cloud training templates are built based on a four-layer decoupling template. Each decoupling template is independently encapsulated and can be updated separately. The four decoupling templates are the basic operation layer, the surveying software layer, the spatial data layer, and the training configuration layer. The basic operation layer is a standardized runtime image, the surveying software layer is a standardized encapsulated pluggable OCI container image plugin, the spatial data layer is a standardized encapsulated surveying training-specific dataset, and the training configuration layer is a set of exclusive configurations for surveying training scenarios. Based on customer requirements, the basic operation layer, surveying software layer, spatial data layer, and training configuration layer are customized to generate personalized cloud training templates.

[0007] Furthermore, the tenant information database is used to store tenant information entered by customers. Tenant information includes identity ID, the personalized cloud training template to which it belongs, permission level and data security access level. When a tenant accesses the cloud through a physical terminal, the cloud authenticates the physical terminal based on the tenant information database. The cloud communicates with the physical terminal that has been authenticated and allocates an independent virtual control plane, data plane and four-dimensional resource quota to the tenant.

[0008] Furthermore, when a tenant sends a training environment call command to the cloud training environment real-time creation module, under the constraints of the virtual control plane, data plane and four-dimensional resource quota allocated by the tenant, the cloud training environment real-time creation module creates a single training environment / training environment cluster for the tenant based on the personalized cloud training template to which the tenant belongs, and establishes a WebRTC remote desktop connection between the physical terminal and the training environment or training environment cluster. The training environment cluster consists of multiple single training environment clusters.

[0009] Furthermore, the process of creating a single training environment / training environment cluster for a tenant based on the tenant's personalized cloud training template includes: Based on the four-layer decoupling template of the personalized cloud training template, the construction process of a single training environment / training environment cluster is divided into basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask. The basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask are executed in parallel.

[0010] Furthermore, the process of obtaining the target computing node in a single training environment includes: Read the GPU adaptation requirements of the pluggable OCI container image plugin in the personalized cloud training template to which the tenant belongs, obtain the computing nodes in the cloud that are fully matched with the GPU adaptation requirements from the node pool in the cloud, build a temporary scheduling pool, and include the computing nodes that are fully matched with the GPU adaptation requirements into the temporary scheduling pool. Determine whether there are any compute nodes in the temporary scheduling pool that have cached the surveying and mapping training exclusive dataset from the personalized cloud training template belonging to the tenant to the local SSD. If there are, retain the compute nodes that have cached the surveying and mapping training exclusive dataset to the local SSD and mark them as first priority. If there are no, determine the availability zone of the compute nodes in the temporary scheduling pool and the storage nodes of the surveying and mapping training exclusive dataset. Mark the compute nodes in the temporary scheduling pool that are marked as availability zones as second priority. First priority is higher than second priority. If the tenant is a single training environment, the computing load of the highest priority computing node in the temporary scheduling pool is extracted. If the computing load of the highest priority computing node has reached the computing load limit of the computing node, the computing load of the second highest priority computing node in the temporary scheduling pool is extracted. The computing node with the lowest computing load among the second highest priority computing nodes is selected as the target computing node for the single training environment. If the computing load of the highest priority computing node does not reach the computing load limit of the computing node, then the computing node with the lowest computing load among the highest priority computing nodes is selected as the target computing node for the single training environment. Schedule the single training environment to the target computing node.

[0011] Furthermore, the process of obtaining the target computing nodes of the training environment cluster includes: If the tenant is a training environment cluster, then node topology labels and node load labels are injected into the computing nodes in the temporary scheduling pool, and business labels are injected into the training environment cluster. The temporary scheduling pool and the training environment cluster were then screened in the first phase, and the steps were as follows: Step s1: Read the tenant ID and cluster ID from the business tag of the single training environment to be scheduled in the training environment cluster, and determine whether there is a single training environment with the same tenant ID and cluster ID that has been scheduled. If so, obtain the node topology tag of the computing node to which the single training environment to be scheduled belongs. If not, proceed directly to the second stage of screening. Step s2: Mark the computing nodes in the temporary scheduling pool that have the same node topology label as the computing nodes belonging to the single training environment as candidate nodes; The temporary scheduling pool and the training environment cluster were then subjected to a second phase of screening, with the following steps: Step q1: Read the class / batch ID from the business tag of the single training environment to be scheduled, and remove nodes from the candidate nodes that already have a single training environment running with the same class / batch ID and nodes that already have a single training environment running with more than or equal to k different tenant IDs; Step q2: Retrieve the node load tags of each candidate node in the temporary scheduling pool, select the candidate node with the lowest load as the target computing node, and schedule the training environment to be scheduled to the target computing node.

[0012] Furthermore, the process of triggering the pre-launch of the personalized cloud training template based on the estimated call timestamp and confidence level includes: The historical call records of several personalized cloud training templates in the templated management module of surveying and mapping training scenarios are used as training data to build a template call prediction model. The template call prediction model is trained using the training data to obtain the completed template call prediction model. Based on the template call prediction model, the predicted call timestamps and confidence scores of several personalized cloud training templates are output. A pre-set confidence threshold is used to pre-start personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is greater than the confidence threshold. Personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is less than or equal to the confidence threshold will not be pre-started. Instead, they will be built in real time when the tenant sends the training environment call command.

[0013] Furthermore, the process of pre-creating a lightweight single training environment / training environment cluster based on a pre-started personalized cloud training template, and performing idle reclamation or rapid activation operations on the lightweight single training environment / training environment cluster includes: Based on the historical call records of the personalized cloud training template, obtain the historical average time consumption for building a single training environment / training environment cluster based on the personalized cloud training template. Based on the estimated call timestamp of the personalized cloud training template and the historical average time consumption of the single training environment / training environment cluster, obtain the pre-creation timestamp of the single training environment / training environment cluster. The target computing nodes for a single training environment / training environment cluster are obtained through the intelligent scheduling module at the pre-created timestamp. Based on the four-layer decoupling template of the personalized cloud training template, the construction process of a single training environment / training environment cluster is divided into lightweight sub-tasks for basic environment scheduling, lightweight sub-tasks for software image loading, lightweight sub-tasks for spatial data mounting, and sub-tasks for training configuration distribution. The lightweight sub-tasks for basic environment scheduling, lightweight sub-tasks for software image loading, lightweight sub-tasks for spatial data mounting, and sub-tasks for training configuration distribution are executed in parallel, thus pre-creating a lightweight single training environment / training environment cluster. Monitor the tenant's training environment call command, preset the idle recycling threshold, and set the idle recycling time stamp based on the estimated call timestamp of the lightweight single training environment / training environment cluster and the idle recycling threshold. If no training environment call command is received from the tenant before the idle recycling time stamp, the lightweight single training environment / training environment cluster will be recycled. If a tenant's training environment call instruction is received before the idle recycling timestamp, a quick activation operation for a lightweight single training environment / training environment cluster is performed, and a WebRTC remote desktop connection is established between the physical terminal and the single training environment / training environment cluster.

[0014] Furthermore, the process of executing the elastic scaling rules of the training configuration layer in the personalized cloud training template to generate elastic operation instructions includes: The rules priority of different elastic scaling rule types are preset. The elastic scaling rule types include training step trigger rules, surveying task trigger rules, resource index trigger rules, time-series pre-trigger rules and permission constraint rules, and the following steps are executed; Step u1: When a single training environment or training environment cluster is created, the elastic scaling rules of the training configuration layer in the corresponding personalized cloud training template are automatically loaded; Step u2: Collect four types of indicator data in real time from a single training environment / training environment cluster, match the four types of indicator data with the elastic scaling rules to determine the trigger conditions, and filter out all triggered elastic scaling rules. Step u3: If multiple elastic scaling rules are triggered simultaneously, obtain the rule priority of each elastic scaling rule, sort them according to the rule priority, and select the elastic scaling rule with the highest rule priority to generate an elastic decision. Step u4: Select the elastic scaling rule with a lower rule priority, and determine whether there is a rule conflict between the elastic scaling rule with the lower rule priority and the elastic decision. If there is no rule conflict, the elastic scaling rule with the lower rule priority is added to the elastic decision. Step u5: Repeat step u4 until no elastic scaling rules are added to the elastic decision, and then convert the elastic decision into elastic operation instructions.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs personalized cloud training templates through a four-layer decoupled template architecture. Each layer is independently encapsulated and can be updated separately. This not only solves the problems of poor reusability and difficulty in adaptation of traditional integrated images, but also supports the rapid adjustment of the basic operating environment, surveying and mapping software combination, spatial dataset and training configuration according to customer needs. It flexibly adapts to the full-scene surveying and mapping training needs such as GIS basic operations, remote sensing image processing, and real-scene 3D modeling.

[0016] Meanwhile, the cloud training environment real-time creation module breaks down environment building into four parallel sub-tasks, which greatly improves the batch building efficiency of single training environments and training environment clusters, and completely gets rid of the time-consuming limitations of traditional solutions.

[0017] 2. The intelligent scheduling module, specifically designed for surveying and mapping training, first filters target computing nodes based on GPU compatibility requirements, then prioritizes them based on dataset caching and availability zone characteristics, ultimately selecting the node with the optimal load. This fundamentally avoids issues such as driver incompatibility, hardware acceleration failure, and high data access latency. For the training environment cluster, dual filtering using node topology tags and business tags ensures efficient node communication within the cluster and cross-tenant load balancing, significantly improving the stability and efficiency of distributed surveying and mapping tasks.

[0018] 3. Optimize the tenant training experience: The template call prediction module builds a prediction model based on historical call data and accurately triggers pre-startup by combining confidence thresholds, avoiding resource waste caused by invalid preloading. The pre-startup execution module creates a lightweight training environment in advance based on the estimated timestamp, completing basic environment configuration and data mounting without starting heavy computing components. Resources are automatically reclaimed when idle, and tenants can quickly activate it when calling, achieving a "second-level entry into the training environment" experience, completely solving the pain point of long environment startup waiting time in traditional solutions. Attached Figure Description

[0019] Figure 1 This is a flowchart of a container orchestration-based dynamic elastic cloud training environment rapid construction system according to an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1 As shown, the system for rapidly building a dynamic elastic cloud training environment based on container orchestration includes a cloud platform. The cloud communication connection includes a template management module for surveying and mapping training scenarios, a tenant information database, a real-time cloud training environment creation module, an intelligent scheduling module, a template call estimation module, a pre-startup execution module, and an elastic scaling rule module. The template management module for surveying and mapping training scenarios is used to build and store personalized cloud training templates based on customer requirements; The real-time cloud training environment creation module is used to create single training environments or training environment clusters for tenants in batches based on the personalized cloud training templates to which the tenant belongs. The intelligent scheduling module is used to perform computing node resource matching and obtain the target computing nodes for a single training environment / training environment cluster during the creation process. The template call prediction module is used to build a template call prediction model, obtain the predicted call timestamp and confidence level of the personalized cloud training template, and trigger the pre-launch of the personalized cloud training template based on the predicted call timestamp and confidence level of the personalized cloud training template. The pre-startup execution module is used to pre-create lightweight single training environments / training environment clusters based on pre-startup personalized cloud training templates, and to perform idle recycling or rapid activation operations on lightweight single training environments / training environment clusters. The elastic scaling rules module is used to execute the elastic scaling rules of the training configuration layer in the personalized cloud training template and generate elastic operation instructions.

[0022] It should be further explained that, in the specific implementation process, the process of building a personalized cloud training template includes: Personalized cloud training templates are built based on a four-layer decoupling template. Each decoupling template is independently encapsulated and can be updated separately. The four decoupling templates are the basic operation layer, the surveying software layer, the spatial data layer, and the training configuration layer. The basic operation layer is a standardized runtime image adapted to different hardware architectures and operating systems. The surveying software layer is a standardized encapsulated pluggable OCI container image plugin. The spatial data layer is a standardized encapsulated dataset dedicated to surveying training. The training configuration layer is a set of dedicated configurations for surveying training scenarios. Among them, the basic operating layer (the bottom layer) is the "operating base" for all upper layers, supporting x86 / arm architecture, domestically produced chips (Phytium, Kunpeng, Hygon, Loongson), and operating systems (Windows Server / Linux / Tongxin UOS / Galaxy Kylin). Built-in driver environment for NVIDIA / AMD / domestic Jingjia Micro JM series GPUs, graphical interface runtime components (X11, Wayland, WebRTC high-definition remote desktop protocol), general dependency library for surveying software, and security hardening strategy; Surveying software layer (middle layer 1): Relies on the basic operation layer, providing a "tool set" for practical training, covering professional software for all surveying scenarios, and solving pain points such as software deployment, version management, and license adaptation; The plugins cover all categories of surveying and mapping software: domestic GIS software (SuperMap series, Aerospace Hongtu PIE-Engine, Zhongdi Digital MapGIS, Wuhan GeoStar), general commercial software (ESRI ArcGIS Pro, ENVI, ERDASIMAGINE), aerial surveying and 3D modeling software (ContextCapture, DJITerra, PhotoScan), point cloud processing software (CloudCompare, LiDAR360), BeiDou data processing software, etc. Each plugin is independently packaged, including the software itself, dependency libraries, GPU acceleration configuration, floating license adaptation script, and version number. It supports visual drag-and-drop combination, eliminating the need to recreate the complete image. Teachers can quickly combine different software training environments. The built-in plugin compatibility verification unit automatically detects port conflicts, dependencies, GPU compatibility, and version compatibility between software to avoid runtime failures after template release. Spatial Data Layer (Intermediate Layer 2): Relies on the basic operation layer to provide "data resources" for practical training, solving the pain points of repeated storage, slow loading, and difficulty in distribution of TB-level spatial data; The datasets are categorized according to training scenarios, including remote sensing image data, digital topographic maps, oblique photogrammetry 3D models, LiDAR point cloud data, BeiDou positioning data, vector space data, and real-scene 3D city models. The architecture adopts a separation of "read-only basic dataset + writable training dataset": the read-only basic dataset is shared globally through distributed shared storage, eliminating the need for repeated copying for each tenant and saving more than 90% of storage resources; the writable training dataset is an encrypted storage volume independently allocated to each tenant, which can only be accessed and modified by the tenant themselves. It has a built-in data compliance verification unit that automatically labels the classification level of datasets (top secret / confidential / secret / public) and matches them with the corresponding confidentiality control rules. It also supports automatic desensitization of surveying and mapping data, and performs coordinate offsetting and accuracy downgrading on high-precision classified data to generate non-classified data that can be used for training.

[0023] Training configuration layer (top layer): Independent of the upper layer, providing "rule-driven" support for the entire layer; The configuration includes: training task steps, computing power elastic scaling rules, spatial data mounting rules, data confidentiality control rules, software license allocation rules, network topology configuration, automatic verification rules for training results, access permission configuration, and environment lifecycle rules. It supports teachers in configuring differentiated computing power, software, and data permissions for different training steps. For example, only low-configuration resources are allocated for basic GIS operation steps, while GPU computing power and corresponding datasets are automatically allocated for 3D modeling steps.

[0024] The next layer provides the operational foundation for the previous layer, and the previous layer depends on the next layer to implement functions. Finally, through the rules of the training configuration layer, a cloud training template that is "made once, combined flexibly, and compatible across environments" is formed.

[0025] Based on customer requirements, the basic operation layer, surveying software layer, spatial data layer, and training configuration layer are customized to generate personalized cloud training templates.

[0026] It needs further explanation that, in the specific implementation process, the tenant information database is used to store tenant information entered by customers. Tenant information includes identity ID, the personalized cloud training template to which they belong, permission level (student / teacher / administrator), and data security access level. When a tenant accesses the cloud through a physical terminal (PC / laptop), the cloud authenticates the physical terminal based on the tenant information database (using a "Single Sign-On (SSO) + JWT token authentication" mechanism, integrating multi-factor authentication (optional SMS / dynamic password), supporting integration with the smart campus platform and third-party teaching systems, enabling access to all bound resources with a single login). The cloud communicates with the physical terminal that has passed tenant authentication and assigns an independent Kubernetes virtual control plane to the tenant (based on vCluster virtual cluster technology, assigning an independent Kubernetes virtual control plane to each tenant (student / group), allowing tenants to complete distributed spatial big data cluster deployment, SuperMapiServer distributed cluster construction, and SP within their own virtual cluster). The system includes cluster-level training operations such as AK spatial analysis cluster configuration. All operations are only effective within the virtual cluster and do not affect the underlying physical cluster or other tenants, solving the core pain point that traditional solutions cannot support production-level cluster training. It also features a data plane (allocating independent network namespaces and encrypted PersistentVolume storage volumes to each tenant's virtual cluster, and using Calico network policies to deny all cross-tenant network access by default, only opening communication ports within the distributed cluster of the tenant according to training needs; for spatial data, it achieves global sharing of read-only datasets and independent encrypted isolation of writable datasets for each tenant, preventing cross-tenant access and tampering of others' data, thus eliminating the risk of data leakage) and four-dimensional resource quotas (configuring independent four-dimensional resource quotas for each tenant and each training template, including hard and soft quotas for CPU core count, memory capacity, GPU memory / shard count, and storage IO bandwidth; hard quotas are the upper limit of resources that a tenant can use, preventing malicious resource preemption; soft quotas are the default quotas for training and can be dynamically adjusted through the elastic engine to ensure stable cluster operation).

[0027] It should be further explained that, in the specific implementation process, when a tenant sends a training environment call command to the cloud training environment real-time creation module, under the constraints of the virtual control plane, data plane, and four-dimensional resource quota allocated by the tenant (the virtual control plane is used to constrain the creation and management permissions of the training environment; the data plane is used to constrain the network and storage isolation of the training environment; and the four-dimensional resource quota is used to constrain the resource usage limit of the training environment), the cloud training environment real-time creation module creates a single training environment / training environment cluster for the tenant based on the personalized cloud training template to which the tenant belongs (in a traditional computer room, each student corresponds to a physical terminal, and all software (such as ArcGIS, ContextCapture), data (remote sensing imagery, point cloud data), and configurations are stored here). The system runs on a physical terminal; in this system, the training environment is the cloud-based digital carrier of this physical terminal—it is an independent and isolated operating environment that contains all the core components needed to support students in completing surveying and mapping training, such as the pluggable OCI plugins of the "surveying and mapping software layer" (such as SuperMap and ContextCapture). Ultimately, these are all deployed in the training environment in the form of containers. The environment provides a stable operating environment for the software (including dependency libraries, GPU drivers, and graphical interface components), ensuring that the software's GPU acceleration, graphics rendering, and other functions work normally. It also establishes a WebRTC remote desktop connection between the physical terminal and the training environment or training environment cluster. The training environment cluster consists of multiple single training environment clusters.

[0028] It should be further explained that, in the specific implementation process, the process of creating a single training environment / training environment cluster for a tenant based on the tenant's personalized cloud training template includes: Based on the four-layer decoupling template of the personalized cloud training template, the construction process of a single training environment / training environment cluster is divided into basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask. The basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask are executed in parallel.

[0029] Furthermore, based on the "read-only layer + writable layer" separation feature of container images, when personalized cloud training templates are updated, only the changed software plugins, configuration files, and dataset shards are updated incrementally, without having to rebuild the entire environment, reducing update time by more than 90%. At the same time, a one-click reset function is provided to address environment corruption and data disorder caused by student misoperation. During the reset, only the writable layer of the container needs to be discarded, and the initial writable layer needs to be remounted, while retaining the base image and read-only dataset. The environment reset can be completed within 3 seconds, restoring it to the initial training state, completely solving the pain point of traditional solutions requiring reconstruction for reset.

[0030] Among them, the basic environment scheduling subtask is used to complete virtual cluster creation, namespace initialization, resource quota configuration, GPU allocation, and network policy distribution in parallel; The software image loading subtask is used to invoke the pluggable OCI container image plugin for surveying software; The spatial data mounting subtask is used to mount standardized and packaged surveying and mapping training datasets. Read-only basic datasets are mounted directly through shared storage; writable training datasets are created in parallel to create encrypted isolated volumes to complete the data mounting. The training configuration and license allocation subtask is used to distribute training configurations, flexible rules, software licenses, experiment manuals, and result verification scripts in parallel.

[0031] It should be further explained that, in the specific implementation process, the process of obtaining the target computing node of a single training environment includes: Read the GPU compatibility requirements of the pluggable OCI container image plugin in the personalized cloud training template of the tenant, and obtain the computing nodes in the cloud that are fully compatible with the GPU compatibility requirements from the node pool (assuming there are 10 computing nodes in total, nodes 1-5 are NVIDIA A10 GPUs, and nodes 6-10 are Jingjia Micro JM9230 GPUs). For example, ContextCapture is best adapted to NVIDIA RTX / A series GPUs, and the domestic version of SuperMap is best adapted to Jingjia Micro JM9230 GPUs. By matching the corresponding cloud GPU nodes, problems such as driver incompatibility and hardware acceleration failure are avoided. A temporary scheduling pool is built, and computing nodes that are fully compatible with the GPU compatibility requirements are included in the temporary scheduling pool. Determine if there are any compute nodes in the temporary scheduling pool that have cached the surveying and mapping training dataset required to be mounted in the personalized cloud training template of the tenant to the local SSD. If so, retain the compute nodes that have cached the surveying and mapping training dataset to the local SSD and mark them as first priority. If not, determine the availability zone of the compute nodes and storage nodes of the surveying and mapping training dataset in the temporary scheduling pool (all compute nodes (GPU servers) and storage nodes (distributed shared storage / MinIO) in the same availability zone must be directly connected through a 10 Gigabit / 25G / InfiniBand high-speed LAN, connected to the same core switch / switch group, with network communication latency between nodes ≤1ms and bandwidth ≥10Gbps to meet the high-speed transmission requirements of TB-level remote sensing images and point cloud data for surveying and mapping, as well as the RPC real-time communication requirements of the distributed cluster). Mark the compute nodes in the temporary scheduling pool that are marked as availability zones as second priority, with first priority being higher than second priority. If the tenant is a single training environment, the computing load of the highest priority computing node in the temporary scheduling pool is extracted. If the computing load of the highest priority computing node has reached the computing load limit of the computing node, the computing load of the second highest priority computing node in the temporary scheduling pool is extracted. The computing node with the lowest computing load among the second highest priority computing nodes is selected as the target computing node for the single training environment. If the computing load of the highest priority computing node does not reach the computing load limit of the computing node, then the computing node with the lowest computing load among the highest priority computing nodes is selected as the target computing node for the single training environment. A single training environment is scheduled to the target compute node. The single training environment is the smallest scheduling atom and has the core attribute of being bound to a single node. A single training environment will only be scheduled to one compute node during its life cycle. All containers, resources (GPUs), and storage mounting are completed within this node. When a single training environment accesses data, it either reads the cache of the local SSD of this node (optimal) or accesses the storage node across availability zones through the network (second best).

[0032] It should be further explained that, in the specific implementation process, the process of obtaining the target computing nodes of the training environment cluster includes: If the tenant is a training environment cluster, then node topology labels and node load labels are injected into the computing nodes in the temporary scheduling pool, and business labels are injected into the training environment cluster. The injected labels are shown in Table 1 below. Table 1 Node topology tags topology.kubernetes.io / zone AZ1 (Availability Zone 1) The physical availability zone that identifies the compute node Node topology tags node.kubernetes.io / switch SW1 (Switch 1) Identifying the physical switch to which the compute node is connected Node load labels node.geo-training / heavy-gpu-count 8 (8 heavy GPUs are already running) Number of recomputational resources on statistical nodes (customized surveying and mapping) Business Tags tenant-id T20240311001 (Tenant's Unique ID) Identify the tenant Business Tags cluster-id cc-cluster-001 (cluster ID) Does the identifier belong to the same training environment cluster? Business Tags task-type heavy-gpu / light / spark-compute Identify the task type (GPU-intensive / Lightweight / Spark compute). Business Tags class-id 2024-ce-1 (Class / Batch ID) Identify the class / batch The first phase of screening is performed on the temporary scheduling pool and the training environment cluster. The steps are as follows: Step s1: Read the tenant ID and cluster ID from the business tag of the single training environment to be scheduled in the training environment cluster, and determine whether there is a single training environment with the same tenant ID and cluster ID that has been scheduled. If there is, obtain the node topology tag of the computing node to which the single training environment to be scheduled belongs. If not (the first single training environment to be scheduled in the training environment cluster), then this stage is not constrained and the second stage of screening is directly carried out. Step s2: Mark the compute nodes in the temporary scheduling pool that have the same node topology label as the compute nodes of the single training environment as candidate nodes (must be in the same zone to avoid high latency across availability zones (usually ≥10ms); preferably in the same switch to further reduce RPC communication latency between clusters (≤1ms)). The temporary scheduling pool and the training environment cluster were then subjected to a second phase of screening, with the following steps: Step q1: Read the class / batch ID from the business tag of the single training environment to be scheduled, and remove nodes that already have a single training environment running with the same class / batch ID from the candidate nodes (prioritize dispersing different classes / batch) and nodes that already have a single training environment running with more than or equal to k (k=3) different tenant IDs (to avoid excessive mixing of node loads). Step q2: Retrieve the node load tags of each candidate node in the temporary scheduling pool, select the candidate node with the lowest load as the target computing node, and schedule the training environment to be scheduled to the target computing node.

[0033] It should be further explained that, in the specific implementation process, the process of building a template call prediction model, obtaining the estimated call timestamp and confidence level of the personalized cloud training template, and triggering the pre-launch of the personalized cloud training template based on the estimated call timestamp and confidence level includes: The template call prediction model is constructed by using historical call records (including the tenant's major (surveying engineering / geographic information science), grade, class ID, batch type (daily training / intensive training week / skills assessment), actual call timestamp (accurate to the minute), call end timestamp, call duration, day of the week, school calendar (start of school / holiday / exam week), and course schedule associated with the smart campus platform (course name, class time, and class)) from the template training scenario template management module as training data. The model is then trained using the training data to obtain the completed template call prediction model. Based on the model, the predicted call timestamps and confidence scores of several personalized cloud training templates are output. The specific process includes: From the historical logs of the templated management module for surveying and mapping training scenarios, several historical call records of personalized cloud training templates were collected as the raw dataset, and feature encoding was performed on the raw dataset: Categorical features (template type, major, grade, day of the week, whether it is a training week): use One-Hot encoding (when the data volume is small) or Embedding encoding (when the data volume is large) to convert them into numerical vectors that the model can process; Time features: Perform periodic encoding (e.g., use sin / cos functions to represent the 24-hour and 7-day cycles to avoid the model treating "23 o'clock" and "0 o'clock" as completely unrelated times), and extract "time period" features (morning 1-4 periods / afternoon 5-8 periods / evening self-study). The "actual timestamp of the template call" is used as the supervision label (GroundTruth) of the model, and in order to adapt to the regression model, the timestamp is converted into "minutes from 0:00 on the same day" (e.g., 8:55 AM on Thursday is converted into 535 minutes). The original dataset is strictly divided according to time order (to avoid data leakage). For example, "historical data from the first 8 weeks" is used as the training set, "data from the 9th week" is used as the validation set, and "data from the 10th week" is used as the test set to simulate a real scenario of using the past to predict the future. Then, feature extraction was performed on the original dataset, including: 1. Template historical statistical characteristics (capturing the template's own calling patterns): Weekly / monthly calling statistics of the template: such as the number of times the template was called on "Day of the Week" in the past 4 weeks, the average calling time, and the standard deviation of the calling time (to determine whether the calling time is stable); Time period preference characteristics of the template: such as whether 90% of the template's calls are concentrated in "morning 1-4 periods" or "evening self-study"; Professional binding characteristics of the template: such as "3D modeling template" is only used by second-year students majoring in surveying and mapping engineering, and "BeiDou data processing template" is only used by students majoring in navigation engineering.

[0034] 2. Tenant group behavior characteristics (capturing class / major training habits): Class historical call habits: such as a class that is accustomed to entering training "10 minutes early", and a class that is accustomed to calling the template only when the class bell rings (8:00); Major training cycle characteristics: such as the 8th-10th week of each semester for the surveying and mapping engineering major is the "intensive training week", during which the call volume of 3D modeling templates will increase by 3 times; Tenant historical operation lag characteristics: such as the last time the class called the template was 5 minutes later than the time on the class schedule, the model can adjust the prediction accordingly.

[0035] 3. External environment auxiliary features (capturing strong constraints of school calendar / timetable): School calendar features: whether it is "the first week of school" (call time may be unstable), whether it is "exam week" (increased calls for self-training), whether it is "holiday" (no calls); Timetable associated features: if the template is associated with a fixed timetable, then extract "timetable class time" as a strong feature, and the model only needs to be fine-tuned based on the timetable (e.g., 5 minutes earlier / later).

[0036] Since geographic mapping template call prediction is a regression problem involving "mixed features (classification + continuous) + time series patterns", the Temporal Fusion Transformer (TFT) model is selected for multivariate time series prediction. It can simultaneously handle "static covariates" (template type, major) and "dynamic covariates" (time features, academic calendar features). Furthermore, TFT has a built-in attention mechanism that can automatically focus on key information such as "call patterns over the past few weeks" and "course schedule times". Model input and output: Input: Static covariates (template ID, major), dynamic covariates (call time of the past 7 days, current academic calendar / time feature); Output: Estimated timestamp (in minutes) of template invocation, and confidence level.

[0037] Training process: The core of this stage is to train the model to learn the rules and to evaluate whether the model meets the requirement of "preloading time error ≤ 10 minutes" through verification.

[0038] Input the preprocessed training set into the model and initialize the model parameters (e.g., LightGBM learning rate = 0.05, tree depth = 6). Forward propagation calculates the difference between the estimated call time and the actual call time (MAE). Backpropagation updates the model parameters, minimizing the error; Repeat the iteration until the validation set error no longer decreases (early stopping mechanism to avoid overfitting). Hyperparameter tuning: Use Bayesian optimization (more efficient than grid search) to tune hyperparameters, with the goal of minimizing the MAE of the validation set.

[0039] Evaluation indicators: Mean Absolute Error (MAE): A core indicator, requiring ≤10 minutes (e.g., if the estimated time is 8:55 and the actual time is 9:00, the error is 5 minutes, which meets the requirement). Maximum absolute error: ≤30 minutes (to avoid extreme errors causing preloading to be too late); A pre-set confidence threshold is used to pre-start personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is greater than the confidence threshold. Personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is less than or equal to the confidence threshold will not be pre-started. Instead, they will be built in real time when the tenant sends the training environment call command.

[0040] It should be further explained that, in the specific implementation process, the process of pre-creating a lightweight single training environment / training environment cluster based on the pre-started personalized cloud training template, and performing idle reclamation or rapid activation operations on the lightweight single training environment / training environment cluster includes: Based on the historical call records of the personalized cloud training template, obtain the historical average time consumption for building a single training environment / training environment cluster based on the personalized cloud training template. Based on the estimated call timestamp of the personalized cloud training template and the historical average time consumption of the single training environment / training environment cluster, obtain the pre-creation timestamp of the single training environment / training environment cluster. At the pre-creation timestamp, the intelligent scheduling module obtains the target computing nodes for a single training environment / training environment cluster (filtering nodes whose GPU models, driver versions, and templates match; prioritizing nodes with pre-cached datasets; scheduling cluster Pods of the same tenant to the same availability zone / same switch; distributing pre-started Pods from different classes to different nodes to avoid single-node overload). Based on the four-layer decoupling template of the personalized cloud training template, the single training environment / training environment cluster construction process is broken down into lightweight sub-tasks for basic environment scheduling (parallel creation of the tenant's vCluster virtual control plane (if not created), namespace, resource quotas (only soft quotas are allocated, hard quotas are reserved), and network policies), and lightweight sub-tasks for software image loading. The system directly calls the pre-cached image layering on the node, creates the container framework of the Pod, but does not start the main process of the surveying software (only starts the basic runtime container and WebRTC remote desktop component)), the lightweight subtask of mounting spatial data (read-only basic dataset is directly mounted through shared storage (no need to copy); the encrypted writable storage volume (PV) of the tenant is created in parallel, but it is not formatted yet (to save time)) and the subtask of distributing training configuration (distributing elastic rules, confidentiality control rules, and experimental manuals in parallel, but not allocating software licenses yet (licenses are dynamically allocated)), the lightweight subtask of basic environment scheduling, the lightweight subtask of loading software image, the lightweight subtask of mounting spatial data, and the subtask of distributing training configuration are executed in parallel, and a lightweight single training environment / training environment cluster is pre-created; Monitor tenants' training environment invocation commands, preset idle reclamation threshold (15min), and set idle reclamation timestamp (estimated invocation timestamp + 15min) based on the estimated invocation timestamp of the lightweight single training environment / training environment cluster and the idle reclamation threshold. If no training environment invocation command is received from the tenant before the idle reclamation timestamp, the lightweight single training environment / training environment cluster is reclaimed, including: releasing GPU / CPU resources; unloading datasets and images; releasing reserved software licenses; and retaining the virtual control plane and encrypted storage volume (the Pod can be quickly rebuilt when the tenant makes subsequent invocations). If a tenant's training environment call instruction is received before the idle recycling timestamp, a fast activation operation for a lightweight single training environment / training environment cluster is performed, including: starting the main process of the surveying software (such as ContextCapture, SuperMap); formatting and mounting the encrypted writable storage volume; dynamically allocating reserved software licenses; adjusting Pod resources to hard quotas (such as allocating a full GPU); and establishing a WebRTC remote desktop connection between the physical terminal and the single training environment / training environment cluster.

[0041] It should be further explained that, in the specific implementation process, the process of executing the elastic scaling rules of the training configuration layer in the personalized cloud training template and generating elastic operation instructions includes: The rules priority of different elastic scaling rule types is preset. The elastic scaling rule types include training step trigger rules, surveying task trigger rules, resource index trigger rules, time series pre-trigger rules and permission constraint rules. The rule priority is sorted from high to low as training step trigger rules > surveying task trigger rules > permission constraint rules > resource index trigger rules > time series pre-trigger rules, and the following steps are executed. Step u1: When a single training environment or training environment cluster is created, the elastic scaling rules of the training configuration layer in the corresponding personalized cloud training template are automatically loaded; Step u2: Collect four types of indicator data in real time from a single training environment / training environment cluster, match the four types of indicator data with the elastic scaling rules to determine the trigger conditions, filter out all triggered elastic scaling rules, and if only one elastic scaling rule is triggered, directly convert the elastic scaling rule into an elastic operation command. Step u3: If multiple elastic scaling rules are triggered simultaneously, obtain the rule priority of each elastic scaling rule, sort them according to the rule priority, and select the elastic scaling rule with the highest rule priority to generate an elastic decision. Step u4: Select the elastic scaling rule with a lower rule priority, and determine whether there is a rule conflict between the elastic scaling rule with the lower rule priority and the elastic decision. If there is no rule conflict, the elastic scaling rule with the lower rule priority is added to the elastic decision. Step u5: Repeat step u4 until no elastic scaling rules are added to the elastic decision. Then, convert the elastic decision into specific elastic operation instructions, such as scaling up GPU shards for vertical Pods, adjusting the number of instances for horizontal Pods, elastically mounting software components, and pre-scaling resource pools.

[0042] It should be further explained that the four types of indicator data for a single training environment / training environment cluster include infrastructure resource indicators (CPU, memory, GPU, video memory, IO bandwidth, cluster resource reserves), surveying and mapping training status indicators (completion of training steps, running status of surveying and mapping tasks, running status of software, and mounting status of datasets), tenant behavior indicators (number of online users, operation type, task submission status, and idle time of the environment), and license status indicators (number of available floating licenses and license utilization rate). The rules are triggered by the following conditions: 1. Training Step Trigger: When students complete a designated training step and move to the next stage, the corresponding flexible operation is automatically executed. For example: Step 1 (GIS Basic Vectorization) is allocated 1 core and 4GB of GPU resources by default; Step 2 (Remote Sensing Image Supervision Classification) is automatically allocated 1 / 4 GPU and 2 cores with 8GB of resources; Step 3 (Oblique Photogrammetry Aerial Triangulation and 3D Modeling) is automatically allocated a full GPU and 8 cores with 32GB of resources. 2. Surveying and mapping task triggering: When students start computationally intensive tasks such as 3D modeling, aerial triangulation, point cloud classification, batch spatial analysis, and distributed image processing, the computing power will be automatically expanded, and the resources will be automatically reduced and released after the task is completed. 3. Resource Metric Trigger: When CPU utilization, memory usage, GPU utilization, VRAM usage, or storage I / O bandwidth reach preset thresholds, automatic scaling operations are executed. For example: if GPU utilization exceeds 80% for 30 seconds, GPU shards / full GPUs are automatically expanded; if it remains below 20% for 5 minutes, GPU resources are automatically reduced to release them. 4. Pre-triggered timing: Based on the course schedule and historical training data, the GPU driver, surveying software image, and training dataset are automatically pre-loaded to the corresponding nodes at a preset time before class, and the GPU resource pool is pre-expanded to avoid insufficient resources during peak class periods.

[0043] 5. License Trigger: Based on the number of available floating licenses, automatically adjust the number of software instances that can run simultaneously to avoid training failure due to insufficient licenses.

[0044] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A system for quickly constructing a dynamic elastic cloud practical training environment based on container orchestration, characterized in that, Including the cloud, the cloud communication connection includes a template management module for surveying and mapping training scenarios, a tenant information database, a real-time creation module for cloud training environments, an intelligent scheduling module, a template call estimation module, a pre-start execution module, and an elastic scaling rule module; The template management module for surveying and mapping training scenarios is used to build and store personalized cloud training templates based on customer requirements; The real-time cloud training environment creation module is used to create single training environments or training environment clusters for tenants in batches based on the personalized cloud training templates to which the tenant belongs. The intelligent scheduling module is used to perform computing node resource matching and obtain the target computing nodes for a single training environment / training environment cluster during the creation process. The template call prediction module is used to build a template call prediction model, obtain the predicted call timestamp and confidence level of the personalized cloud training template, and trigger the pre-launch of the personalized cloud training template based on the predicted call timestamp and confidence level of the personalized cloud training template. The pre-startup execution module is used to pre-create lightweight single training environments / training environment clusters based on pre-startup personalized cloud training templates, and to perform idle recycling or rapid activation operations on lightweight single training environments / training environment clusters. The elastic scaling rules module is used to execute the elastic scaling rules of the training configuration layer in the personalized cloud training template and generate elastic operation instructions.

2. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration as described in claim 1, characterized in that, The process of building a personalized cloud training template includes: Personalized cloud training templates are built based on a four-layer decoupling template. Each decoupling template is independently encapsulated and can be updated separately. The four decoupling templates are the basic operation layer, the surveying software layer, the spatial data layer, and the training configuration layer. The basic operation layer is a standardized runtime image, the surveying software layer is a standardized encapsulated pluggable OCI container image plugin, the spatial data layer is a standardized encapsulated surveying training-specific dataset, and the training configuration layer is a set of exclusive configurations for surveying training scenarios. Based on customer requirements, the basic operation layer, surveying software layer, spatial data layer, and training configuration layer are customized to generate personalized cloud training templates.

3. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 2, characterized in that, The tenant information database is used to store tenant information entered by customers. Tenant information includes identity ID, personalized cloud training template, permission level and data security access level. When a tenant accesses the cloud through a physical terminal, the cloud authenticates the physical terminal based on the tenant information database. The cloud communicates with the physical terminal that has been authenticated and allocates independent virtual control plane, data plane and four-dimensional resource quotas to the tenant.

4. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 3, characterized in that, When a tenant sends a training environment call command to the cloud training environment real-time creation module, under the constraints of the virtual control plane, data plane and four-dimensional resource quota allocated by the tenant, the cloud training environment real-time creation module creates a single training environment / training environment cluster for the tenant based on the personalized cloud training template to which the tenant belongs, and establishes a WebRTC remote desktop connection between the physical terminal and the training environment or training environment cluster. The training environment cluster consists of multiple single training environment clusters.

5. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 4, characterized in that, The process of creating a single training environment / training environment cluster for a tenant based on the personalized cloud training template to which the tenant belongs includes: Based on the four-layer decoupling template of the personalized cloud training template, the construction process of a single training environment / training environment cluster is divided into basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask. The basic environment scheduling subtask, software image loading subtask, spatial data mounting subtask, and training configuration and license allocation subtask are executed in parallel.

6. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 5, characterized in that, The process of obtaining the target computing node for a single training environment includes: Read the GPU adaptation requirements of the pluggable OCI container image plugin in the personalized cloud training template to which the tenant belongs, obtain the computing nodes in the cloud that are fully matched with the GPU adaptation requirements from the node pool in the cloud, build a temporary scheduling pool, and include the computing nodes that are fully matched with the GPU adaptation requirements into the temporary scheduling pool. Determine whether there are any compute nodes in the temporary scheduling pool that have cached the surveying and mapping training exclusive dataset from the personalized cloud training template belonging to the tenant to the local SSD. If there are, retain the compute nodes that have cached the surveying and mapping training exclusive dataset to the local SSD and mark them as first priority. If there are no, determine the availability zone of the compute nodes in the temporary scheduling pool and the storage nodes of the surveying and mapping training exclusive dataset. Mark the compute nodes in the temporary scheduling pool that are marked as availability zones as second priority. First priority is higher than second priority. If the tenant is a single training environment, the computing load of the highest priority computing node in the temporary scheduling pool is extracted. If the computing load of the highest priority computing node has reached the computing load limit of the computing node, the computing load of the second highest priority computing node in the temporary scheduling pool is extracted. The computing node with the lowest computing load among the second highest priority computing nodes is selected as the target computing node for the single training environment. If the computing load of the highest priority computing node does not reach the computing load limit of the computing node, then the computing node with the lowest computing load among the highest priority computing nodes is selected as the target computing node for the single training environment. Schedule the single training environment to the target computing node.

7. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 6, characterized in that, The process of obtaining the target computing node of the training environment cluster includes: If the tenant is a training environment cluster, then node topology labels and node load labels are injected into the computing nodes in the temporary scheduling pool, and business labels are injected into the training environment cluster. The temporary scheduling pool and the training environment cluster were then screened in the first phase, and the steps were as follows: Step s1: Read the tenant ID and cluster ID from the business tag of the single training environment to be scheduled in the training environment cluster, and determine whether there is a single training environment with the same tenant ID and cluster ID that has been scheduled. If so, obtain the node topology tag of the computing node to which the single training environment to be scheduled belongs. If not, proceed directly to the second stage of screening. Step s2: Mark the computing nodes in the temporary scheduling pool that have the same node topology label as the computing nodes belonging to the single training environment as candidate nodes; The temporary scheduling pool and the training environment cluster were then subjected to a second phase of screening, with the following steps: Step q1: Read the class / batch ID from the business tag of the single training environment to be scheduled, and remove nodes from the candidate nodes that already have a single training environment running with the same class / batch ID, as well as nodes that already have a single training environment running with more than or equal to k different tenant IDs; Step q2: Retrieve the node load tags of each candidate node in the temporary scheduling pool, select the candidate node with the lowest load as the target computing node, and schedule the training environment to be scheduled to the target computing node.

8. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 7, characterized in that, The process of triggering the pre-launch of a personalized cloud training template based on the estimated call timestamp and confidence level includes: The historical call records of several personalized cloud training templates in the templated management module of surveying and mapping training scenarios are used as training data to build a template call prediction model. The template call prediction model is trained using the training data to obtain the completed template call prediction model. Based on the template call prediction model, the predicted call timestamps and confidence scores of several personalized cloud training templates are output. A pre-set confidence threshold is used to pre-start personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is greater than the confidence threshold. Personalized cloud training templates whose confidence level for the corresponding estimated call timestamp is less than or equal to the confidence threshold will not be pre-started. Instead, they will be built in real time when the tenant sends the training environment call command.

9. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 8, characterized in that, The process of pre-creating a lightweight single training environment / training environment cluster based on a pre-started personalized cloud training template, and performing idle reclamation or rapid activation operations on the lightweight single training environment / training environment cluster includes: Based on the historical call records of the personalized cloud training template, obtain the historical average time consumption for building a single training environment / training environment cluster based on the personalized cloud training template. Based on the estimated call timestamp of the personalized cloud training template and the historical average time consumption of a single training environment / training environment cluster, obtain the pre-creation timestamp of the single training environment / training environment cluster. At the pre-creation timestamp, the target computing nodes of a single training environment / training environment cluster are obtained through the intelligent scheduling module. Based on the four-layer decoupling template of the personalized cloud training template, the construction process of a single training environment / training environment cluster is divided into lightweight sub-tasks for basic environment scheduling, lightweight sub-tasks for software image loading, lightweight sub-tasks for spatial data mounting, and sub-tasks for training configuration distribution. The lightweight sub-tasks for basic environment scheduling, lightweight sub-tasks for software image loading, lightweight sub-tasks for spatial data mounting, and sub-tasks for training configuration distribution are executed in parallel, thus pre-creating a lightweight single training environment / training environment cluster. Monitor the tenant's training environment call command, preset the idle recycling threshold, and set the idle recycling time stamp based on the estimated call timestamp of the lightweight single training environment / training environment cluster and the idle recycling threshold. If no training environment call command is received from the tenant before the idle recycling time stamp, the lightweight single training environment / training environment cluster will be recycled. If a tenant's training environment call instruction is received before the idle recycling timestamp, a quick activation operation for a lightweight single training environment / training environment cluster is performed, and a WebRTC remote desktop connection is established between the physical terminal and the single training environment / training environment cluster.

10. The rapid construction system for a dynamic elastic cloud training environment based on container orchestration according to claim 9, characterized in that, The process of executing the elastic scaling rules of the training configuration layer in the personalized cloud training template and generating elastic operation instructions includes: The rules priority of different elastic scaling rule types are preset. The elastic scaling rule types include training step trigger rules, surveying task trigger rules, resource index trigger rules, time-series pre-trigger rules and permission constraint rules, and the following steps are executed; Step u1: When a single training environment or training environment cluster is created, the elastic scaling rules of the training configuration layer in the corresponding personalized cloud training template are automatically loaded; Step u2: Collect four types of indicator data in real time from a single training environment / training environment cluster, match the four types of indicator data with the elastic scaling rules to determine the trigger conditions, and filter out all triggered elastic scaling rules. Step u3: If multiple elastic scaling rules are triggered simultaneously, obtain the rule priority of each elastic scaling rule, sort them according to the rule priority, and select the elastic scaling rule with the highest rule priority to generate an elastic decision. Step u4: Select the elastic scaling rule with a lower rule priority, and determine whether there is a rule conflict between the elastic scaling rule with the lower rule priority and the elastic decision. If there is no rule conflict, the elastic scaling rule with the lower rule priority is added to the elastic decision. Step u5: Repeat step u4 until no elastic scaling rules are added to the elastic decision, and then convert the elastic decision into elastic operation instructions.

Images