K8s-based intelligent AI platform resource management and scheduling system

Abstract

The invention belongs to the technical field of resource management and scheduling, and particularly relates to an intelligent AI platform resource management and scheduling system based on K8s. Comprising a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security enhanced mirror image warehouse and a reasoning early warning system. Through the synergistic effect of the modules, the technical problems of inflexible resource scheduling, insufficient resource monitoring, incomplete user resource isolation, limited computing power distribution, complex mirror image management and incomplete operation and maintenance management of an AI platform in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the field of resource management and scheduling technology, and in particular relates to an intelligent AI platform resource management and scheduling system based on K8s. K8s is Kubernetes, that is, container orchestration management technology. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, the demand for computing resources for the development, training, and deployment of AI models is increasing. Deep learning model training, in particular, typically requires a significant amount of computing resources, such as CPUs, NPUs, and GPUs. Furthermore, existing AI platforms commonly suffer from inflexible resource scheduling, incomplete resource monitoring, imperfect user resource isolation, and limited computing resource allocation. These issues are particularly acute in multi-user, multi-tasking scenarios.

[0003] Existing AI platform solutions generally rely on static resource allocation mechanisms, where resources are allocated in fixed increments before tasks are run. This approach lacks flexibility in practical applications and struggles to cope with fluctuations in resource demand between tasks. Furthermore, resource monitoring on existing platforms is primarily focused on a single server or node, making it impossible to comprehensively monitor resource usage by user groups and their internal users, impacting the platform's resource management efficiency.

[0004] In general, existing AI platform solutions include the following technical issues:

[0005] Inflexible resource scheduling: Traditional AI platforms are unable to flexibly adjust resources, resulting in low resource utilization and prone to resource contention or waste.

[0006] Insufficient resource monitoring: It is impossible to monitor user resource usage in real time, management transparency is low, and reasonable scheduling and optimization are difficult.

[0007] Incomplete user resource isolation: When multiple users share the platform, there are data security risks, insufficient resource isolation, and it is difficult to ensure privacy.

[0008] Limited computing power allocation: Computing power resource allocation cannot meet the needs of different tasks, making it difficult to fully utilize resources such as CPU, GPU / NPU.

[0009] Complex image management: Development environment configuration is cumbersome, image sharing and storage mechanisms are imperfect, and user operations are complex.

[0010] Imperfect operation and maintenance management: The lack of real-time monitoring and early warning mechanisms makes it difficult to respond to system failures in a timely manner, affecting platform stability.

[0011] Therefore, how to improve the existing AI platform to solve its existing problems such as inflexible resource scheduling, insufficient resource monitoring, imperfect user resource isolation, limited computing power allocation, complex image management, and imperfect operation and maintenance management is a technical problem that needs to be solved urgently. Summary of the Invention

[0012] The purpose of the present invention is to provide an intelligent AI platform resource management and scheduling system based on K8s to solve the technical problems of existing AI platforms, such as inflexible resource scheduling, insufficient resource monitoring, imperfect user resource isolation, limited computing power allocation, complex image management, and imperfect operation and maintenance management.

[0013] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0014] A K8s-based intelligent AI platform resource management and scheduling system, including a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security-enhanced image repository, and an inference warning system. The dynamic resource prediction module is connected to the K8s cluster management and resource scheduling module, which is connected to the resource monitoring and analysis module, which is connected to the resource isolation and sharing module. The computing power resource allocation module, the security-enhanced image repository, and the inference warning system are provided on a development platform.

[0015] The dynamic resource prediction module uses a multi-granularity LSTM neural network prediction model to generate a four-dimensional resource demand map including CPU / GPU / memory / network bandwidth in real time, and compares and analyzes it with historical load patterns;

[0016] The K8s cluster management and resource scheduling module generates a scheduling plan based on a multi-dimensional priority evaluation model with dynamic weights and implements feedback-based real-time resource scheduling.

[0017] The resource monitoring and analysis module monitors the resource usage of the user group and its members in real time, such as the occupancy of CPU, memory, and computing power resources, and analyzes the scheduling plan;

[0018] The resource isolation and sharing module achieves multi-level resource isolation, GPU memory fragmentation and reorganization, and cross-tenant computing power sharing through hardware-level virtualization segmentation and namespace strategy linkage;

[0019] The computing power resource allocation module automatically adjusts the computing power allocation strategy when the container is running based on a dynamic power consumption-performance balance algorithm driven by reinforcement learning;

[0020] The security-enhanced image repository has a built-in blockchain-based image traceability system, which supports vulnerability scanning and signature verification under a zero-trust architecture.

[0021] The reasoning warning system generates a fault prediction report with confidence rating through time series anomaly detection and Bayesian network root cause analysis.

[0022] Preferably, the dynamic resource prediction module includes a heterogeneous resource association analysis unit, a burst load detection unit and a prediction self-correction unit;

[0023] The heterogeneous resource association analysis unit establishes a nonlinear regression model between GPU memory usage and CUDA core occupancy;

[0024] The burst load detection unit uses an improved isolation forest algorithm to identify abnormal resource demand spikes;

[0025] The prediction self-correction unit automatically triggers online fine-tuning of LSTM network parameters when the actual usage deviates from the predicted value by more than a threshold.

[0026] Preferably, the K8s cluster management and resource scheduling module includes a topology-aware scheduling module, and is provided with an elastic preemption mechanism and an energy-sensitive scheduling strategy;

[0027] The topology-aware scheduling module optimizes the affinity deployment of multiple containers for AI training tasks based on the connection status between nodes;

[0028] The elastic preemption mechanism is a dynamic recycling method for low-priority task containers to retain checkpoints and release resources when receiving high-priority requests;

[0029] The energy consumption-sensitive scheduling strategy dynamically adjusts the node selection weight in combination with the PUE value of the data center.

[0030] Preferably, the execution process of the topology-aware scheduling module is as follows:

[0031] Automatically identify AllReduce operation mode based on NCCL communication library features;

[0032] Prioritize assigning tasks of the parameter server architecture to node groups with RDMA network connections;

[0033] Dynamically adjust the ratio of compute nodes to parameter nodes in mixed precision training scenarios.

[0034] Preferably, the resource isolation and sharing module includes a video memory dynamic reorganization module, a security isolation enhancement layer and a sharing conflict arbitration module;

[0035] The memory dynamic reorganization module implements cross-container time-sharing multiplexing of GPU memory through virtual address remapping;

[0036] The security isolation enhancement layer injects an eBPF program to monitor the CUDA API call path when the container is running;

[0037] The sharing conflict arbitration module dynamically allocates computing time periods based on an auction mechanism when multiple tasks compete for GPU resources.

[0038] Preferably, the computing resource allocation module is based on a dynamic power consumption-performance balance algorithm driven by reinforcement learning, and the execution process of automatically adjusting the computing power allocation strategy when the container is running is as follows:

[0039] Real-time collection of container dynamic environment data and modeling:

[0040] Real-time collection of container-level computing power indicators, including GPU utilization, video memory occupancy, and CUDA core activity, to build a multi-dimensional resource status vector;

[0041] Establishing a power consumption-performance correlation model, and quantitatively calculating the training throughput corresponding to each watt of power consumption through the power consumption-performance correlation model;

[0042] Predict task load fluctuation trends through time series analysis to form a dynamic decision space;

[0043] Building a reinforcement learning decision framework:

[0044] Design a triple model of state space, action space, and reward function. The state space includes cluster resource status, the action space includes computing power allocation strategy, and the reward function includes energy efficiency ratio and task completion rate.

[0045] Use a deep deterministic policy gradient algorithm to generate computing power allocation plans in the continuous action space;

[0046] Deploy an online learning mechanism to dynamically update policy network parameters based on scheduling results;

[0047] Perform real-time strategy execution and feedback:

[0048] Dynamically adjust container resource limits through the Kubelet extension component to achieve millisecond-level response for GPU computing power allocation;

[0049] Monitor the actual energy efficiency ratio after the strategy is executed, and mark the results that deviate from the expected value by more than 5% as negative samples;

[0050] Trigger the resource recovery mechanism to immediately reallocate resources occupied by inefficient allocation strategies.

[0051] Security constraints and exception handling:

[0052] Set hard boundary conditions (such as temperature thresholds and SLA response times) to limit the scope of policy exploration;

[0053] When node-level resource contention is detected, it automatically downgrades to a preset conservative allocation strategy;

[0054] Record historical data of policy iterations for model interpretability analysis and audit tracing.

[0055] Preferably, the security-enhanced image repository includes a blockchain evidence storage and traceability module, a zero-trust vulnerability scanning module, and a factor signature verification module;

[0056] The blockchain evidence storage and traceability module implements Merkle tree structure encryption on the image layer metadata, generates a unique hash value for each layer and anchors it to the consortium chain, records the hash values ​​of all dependent components from the base image to the final build, and forms an unalterable build history track. When the image is pulled, the on-chain verification process is automatically triggered to compare the consistency of the local image hash with the on-chain evidence.

[0057] The zero-trust vulnerability scanning module dynamically generates customized vulnerability scanning rules based on the characteristics of the container operating environment (such as kernel version and dependency library list), uses graph neural networks to identify vulnerability propagation paths between image layers, predicts potential attack surfaces, and automatically triggers sandbox environment testing for high-risk images to capture hidden vulnerabilities;

[0058] The signature verification module automates the generation, rotation, and revocation of signature keys through the HSM hardware security module, embeds K8s cluster fingerprint information when signing the image, ensures that only specific clusters can decrypt and run, and synchronously verifies the image signature and node identity certificate when the container starts, blocking man-in-the-middle attacks.

[0059] Preferably, the inference warning system collects container-level inference latency, GPU memory occupancy, and CUDA core utilization indicators in real time through a multi-dimensional time series data collection and preprocessing unit, constructs a multi-dimensional time series data stream, and performs noise reduction and smoothing on the original data through a sliding window mechanism to eliminate instantaneous interference noise;

[0060] A dynamic time warping algorithm is used to align timestamps of heterogeneous data sources, ensuring synchronous analysis of multi-dimensional data. A composite anomaly detection model is used for statistical baseline detection, and a Gaussian mixture model is established to characterize the normal fluctuation range of each indicator, marking anomalies that deviate from the 3σ threshold. Deep learning is used for early warning, using a time-domain convolutional network (TCN) to capture long-term pattern anomalies and identify gradual faults that are difficult to detect with traditional methods.

[0061] The correlation analysis engine calculates the Pearson correlation coefficient between abnormal indicators and screens out abnormal clusters with strong correlations. Dynamic Bayesian root cause inference is established, and a three-layer Bayesian network is constructed, including hardware layer, framework layer, and application layer nodes. Conditional probability tables are trained based on historical fault data to quantify the causal influence strength between nodes. When an anomaly is detected, the maximum a posteriori probability path is traced back through the variable elimination algorithm to locate the most likely root cause node.

[0062] Confidence levels are rated and reports are generated. Monte Carlo sampling is used to calculate the confidence interval of the root cause inference results, and the confidence levels are divided into high, medium, and low levels according to the probability values. A structured report is generated, including an abnormal timeline map, root cause path derivation, and disposal suggestions. Early warning information is pushed to the operation and maintenance platform in real time through the API gateway, triggering automatic repair or manual intervention processes.

[0063] The beneficial effects of the present invention include:

[0064] The K8s-based intelligent AI platform resource management and scheduling system provided by the present invention includes a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security-enhanced image repository and an inference warning system. Through the synergistic effect of the above modules, it effectively solves the technical problems of inflexible resource scheduling, insufficient resource monitoring, imperfect user resource isolation, limited computing power allocation, complex image management and imperfect operation and maintenance management existing in the AI ​​platform in the prior art.

[0065] Through K8s cluster management and intelligent resource scheduling, the problem of inflexible resource scheduling in traditional AI platforms is solved. Through a multi-level resource isolation mechanism, resource isolation within user groups and between users, as well as a sharing mechanism between user groups, is achieved, ensuring data security and privacy. Detailed resource usage reports are generated through real-time resource monitoring and analysis, improving the transparency of platform management. Computing resource allocation supports flexible allocation of CPU, memory, and computing resources by user group to maximize resource utilization. Image repository management and custom upload functions allow users to upload images of development environments to personal or public image repositories, support image isolation and sharing, monitor server health status through intelligent operation and maintenance monitoring and early warning mechanisms, provide real-time warnings, and generate operation and maintenance reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a schematic diagram of the composition architecture of the K8s-based intelligent AI platform resource management and scheduling system of the present invention. DETAILED DESCRIPTION

[0067] The following is combined with Figure 1 The present invention is described in further detail:

[0068] Example 1

[0069] See attached Figure 1 As shown, a K8s-based intelligent AI platform resource management and scheduling system includes a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security-enhanced image warehouse and an inference warning system. The dynamic resource prediction module is connected to the K8s cluster management and resource scheduling module, the K8s cluster management and resource scheduling module is connected to the resource monitoring and analysis module, the resource monitoring and analysis module is connected to the resource isolation and sharing module, and the computing power resource allocation module, the security-enhanced image warehouse and the inference warning system are arranged on the development platform.

[0070] The dynamic resource prediction module uses a multi-granularity LSTM neural network prediction model to generate a four-dimensional resource demand map encompassing CPU, GPU, memory, and network bandwidth in real time, and compares and analyzes it against historical load patterns. The Kubernetes cluster management and resource scheduling module generates scheduling plans based on a dynamic weighted multi-dimensional priority evaluation model and implements feedback-based real-time resource scheduling. The resource monitoring and analysis module monitors resource usage by user groups and their members in real time, such as CPU, memory, and computing power usage, and analyzes the scheduling plans. The resource isolation and sharing module implements multi-level resource isolation, GPU memory fragmentation and reorganization, and cross-tenant computing power sharing through hardware-level virtualization segmentation and namespace policy linkage. The computing power resource allocation module automatically adjusts computing power allocation policies during container runtime based on a dynamic power-performance balancing algorithm driven by reinforcement learning. The security-enhanced image repository, with its built-in blockchain-based image traceability system, supports vulnerability scanning and signature verification in a zero-trust architecture. The inference warning system generates fault prediction reports with confidence ratings through time series anomaly detection and Bayesian network root cause analysis.

[0071] In this embodiment, the dynamic resource prediction module includes a heterogeneous resource association analysis unit, a burst load detection unit, and a prediction self-correction unit. The heterogeneous resource association analysis unit establishes a nonlinear regression model between GPU memory usage and CUDA core occupancy. The burst load detection unit uses an improved isolation forest algorithm to identify abnormal resource demand spikes. The prediction self-correction unit automatically triggers online fine-tuning of LSTM network parameters when actual usage deviates from the predicted value by more than a threshold.

[0072] Example 2

[0073] Based on Example 1, the K8s cluster management and resource scheduling module includes a topology-aware scheduling module and implements an elastic preemption mechanism and an energy-sensitive scheduling strategy. The topology-aware scheduling module optimizes the affinity deployment of multiple containers for AI training tasks based on the connectivity between nodes. The elastic preemption mechanism is a dynamic recycling method that allows low-priority task containers to retain checkpoints and release resources when they receive high-priority requests. The energy-sensitive scheduling strategy dynamically adjusts node selection weights based on the data center's PUE value.

[0074] The execution process of the topology-aware scheduling module is as follows:

[0075] Automatically identify AllReduce operation mode based on NCCL communication library features;

[0076] Prioritize assigning tasks of the parameter server architecture to node groups with RDMA network connections;

[0077] Dynamically adjust the ratio of compute nodes to parameter nodes in mixed precision training scenarios.

[0078] The resource isolation and sharing module includes a dynamic memory reorganization module, a security isolation enhancement layer, and a sharing conflict arbitration module. The dynamic memory reorganization module implements cross-container time-sharing multiplexing of GPU memory through virtual address remapping. The security isolation enhancement layer injects an eBPF program during container runtime to monitor the CUDA API call path. The sharing conflict arbitration module dynamically allocates computing time slots based on an auction mechanism when multiple tasks compete for GPU resources.

[0079] The computing resource allocation module is based on a dynamic power consumption-performance balance algorithm driven by reinforcement learning. The execution process of automatically adjusting the computing resource allocation strategy when the container is running is as follows:

[0080] Real-time collection of container dynamic environment data and modeling:

[0081] Collect container-level computing power metrics in real time, including GPU utilization, video memory occupancy, and CUDA core activity, to construct a multi-dimensional resource state vector. Establish a power consumption-performance correlation model, using which the training throughput per watt of power consumption is quantified. Predict task load fluctuation trends through time series analysis to form a dynamic decision space. Establish a reinforcement learning decision framework: Design a triple model of state space, action space, and reward function. The state space includes cluster resource status, the action space includes computing power allocation strategies, and the reward function includes energy efficiency ratio and task completion rate. Use a deep deterministic policy gradient algorithm to generate computing power allocation plans in the continuous action space. Deploy an online learning mechanism to dynamically update policy network parameters based on scheduling results.

[0082] Real-time policy execution and feedback: Dynamically adjust container resource limits through the Kubelet extension component to achieve millisecond-level response in GPU computing power allocation; monitor the actual energy efficiency ratio after policy execution, and mark results that deviate from the expected value by more than 5% as negative samples; trigger the resource recovery mechanism to immediately reallocate resources occupied by inefficient allocation strategies.

[0083] Set safety constraints and exception handling: Set hard boundary conditions (such as temperature thresholds and SLA response times) to limit the scope of policy exploration; automatically downgrade to the preset conservative allocation strategy when node-level resource contention is detected; record policy iteration history data for model interpretability analysis and audit traceability.

[0084] During the actual operation of the computing resource allocation module, container-level metrics are first captured in real time. Moving averages of these metrics are calculated using a sliding window. Min-Max normalization is performed on heterogeneous metrics to eliminate dimensional differences. Then, power consumption and performance correlation modeling is performed, an energy efficiency ratio metric is defined, and a polynomial regression model is established based on the energy efficiency ratio metric. The specific formula for the energy efficiency ratio metric is as follows:

[0085] η = training throughput / power consumption × 100%.

[0086] Then, load trend prediction is performed, and the ARIMA model is used to predict task load fluctuations within a specified time in the future, and dynamic decision space boundary conditions are set. The dynamic decision space boundary conditions can be set to the allowed GPU utilization adjustment range and the maximum instantaneous power consumption change rate.

[0087] When establishing a reinforcement learning decision framework, we build a state space, including cluster-level information such as remaining node memory and GPU temperature curve slope, and container-level information such as current energy efficiency, remaining task duration, and SLA achievement rate. We define the action space, GPU computing power allocation ratio, memory limit adjustment step size, and CUDA stream processor priority weight. The reward function is set to:

[0088] R t =α·η t +β·SLA_score-γ·Δp overshoot ;

[0089] α, β, and γ are experimental optimization coefficients.

[0090] The training process of the reinforcement learning decision framework is as follows: the network structure is an actor network (4-layer MLP) + a critic network (3-layer MLP), exploring noise, and the online update mechanism synchronizes the target network parameters every 50 decision cycles. The policy network outputs an action vector, and the device plugin is used to modify container resource limits. Response latency measurements show an average of 120ms from decision generation to resource validation.

[0091] Example 3

[0092] Based on Example 1 or Example 2, the security-enhanced image warehouse includes a blockchain evidence storage and traceability module, a zero-trust vulnerability scanning module, and a factor signature verification module. The blockchain evidence storage and traceability module implements Merkle tree structure encryption on the image layer metadata, generates a unique hash value for each layer and anchors it to the alliance chain, records the hash values ​​of all dependent components from the base image to the final build, forms an unalterable construction history track, automatically triggers the chain verification process when the image is pulled, and compares the consistency of the local image hash with the chain evidence. The zero-trust vulnerability scanning module dynamically generates customized vulnerability scanning rules based on the characteristics of the container operating environment (such as kernel version, dependency library list), uses graph neural networks to identify vulnerability propagation paths between image layers, predicts potential attack surfaces, and automatically triggers sandbox environment operation tests for high-risk images to capture hidden vulnerabilities. The signature verification module realizes the automation of signature key generation, rotation and revocation through the HSM hardware security module, embeds K8s cluster fingerprint information when signing the image, ensures that only specific clusters can decrypt and run, and synchronously verifies the image signature and node identity certificate when the container is started to block man-in-the-middle attacks.

[0093] The inference warning system collects container-level inference latency, GPU memory occupancy, and CUDA core utilization indicators in real time through a multi-dimensional time series data collection and preprocessing unit, constructs a multi-dimensional time series data stream, and performs noise reduction and smoothing on the original data through a sliding window mechanism to eliminate instantaneous interference noise;

[0094] A dynamic time warping algorithm is used to align timestamps of heterogeneous data sources, ensuring synchronous analysis of multi-dimensional data. A composite anomaly detection model is used for statistical baseline detection, and a Gaussian mixture model is established to characterize the normal fluctuation range of each indicator, marking anomalies that deviate from the 3σ threshold. Deep learning is used for early warning, using a time-domain convolutional network (TCN) to capture long-term pattern anomalies and identify gradual faults that are difficult to detect with traditional methods.

[0095] The correlation analysis engine calculates the Pearson correlation coefficient between abnormal indicators and screens out abnormal clusters with strong correlations. Dynamic Bayesian root cause inference is established, and a three-layer Bayesian network is constructed, including hardware layer, framework layer, and application layer nodes. Conditional probability tables are trained based on historical fault data to quantify the causal influence strength between nodes. When an anomaly is detected, the maximum a posteriori probability path is traced back through the variable elimination algorithm to locate the most likely root cause node.

[0096] Confidence levels are rated and reports are generated. Monte Carlo sampling is used to calculate the confidence interval of the root cause inference results, and the confidence levels are divided into high, medium, and low levels according to the probability values. A structured report is generated, including an abnormal timeline map, root cause path derivation, and disposal suggestions. Early warning information is pushed to the operation and maintenance platform in real time through the API gateway, triggering automatic repair or manual intervention processes.

[0097] In summary, the K8s-based intelligent AI platform resource management and scheduling system provided by the present invention includes a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security-enhanced image warehouse and an inference warning system. Through the synergistic effect of the above modules, it effectively solves the technical problems of inflexible resource scheduling, insufficient resource monitoring, imperfect user resource isolation, limited computing power allocation, complex image management and imperfect operation and maintenance management existing in the AI ​​platform in the prior art.

[0098] Through K8s cluster management and intelligent resource scheduling, the problem of inflexible resource scheduling in traditional AI platforms is solved. Through a multi-level resource isolation mechanism, resource isolation within user groups and between users, as well as a sharing mechanism between user groups, is achieved, ensuring data security and privacy. Detailed resource usage reports are generated through real-time resource monitoring and analysis, improving the transparency of platform management. Computing resource allocation supports flexible allocation of CPU, memory, and computing resources by user group to maximize resource utilization. Image repository management and custom upload functions allow users to upload images of development environments to personal or public image repositories, support image isolation and sharing, monitor server health status through intelligent operation and maintenance monitoring and early warning mechanisms, provide real-time warnings, and generate operation and maintenance reports.

Claims

1. A K8s-based intelligent AI platform resource management and scheduling system, characterized by: It includes a dynamic resource prediction module, a K8s cluster management and resource scheduling module, a resource monitoring and analysis module, a resource isolation and sharing module, a computing power resource allocation module, a security-enhanced image warehouse and an inference warning system. The dynamic resource prediction module is connected to the K8s cluster management and resource scheduling module, the K8s cluster management and resource scheduling module is connected to the resource monitoring and analysis module, the resource monitoring and analysis module is connected to the resource isolation and sharing module, and the computing power resource allocation module, the security-enhanced image warehouse and the inference warning system are set on the development platform; The dynamic resource prediction module uses a multi-granularity LSTM neural network prediction model to generate a four-dimensional resource demand map including CPU / GPU / memory / network bandwidth in real time, and compares and analyzes it with historical load patterns; The K8s cluster management and resource scheduling module generates a scheduling plan based on a multi-dimensional priority evaluation model with dynamic weights and implements feedback-based real-time resource scheduling. The resource monitoring and analysis module monitors the resource usage of the user group and its members in real time, such as the occupancy of CPU, memory, and computing power resources, and analyzes the scheduling plan; The resource isolation and sharing module achieves multi-level resource isolation, GPU memory fragmentation and reorganization, and cross-tenant computing power sharing through hardware-level virtualization segmentation and namespace strategy linkage; The computing power resource allocation module automatically adjusts the computing power allocation strategy when the container is running based on a dynamic power consumption-performance balance algorithm driven by reinforcement learning; The security-enhanced image repository has a built-in blockchain-based image traceability system, which supports vulnerability scanning and signature verification under a zero-trust architecture. The reasoning warning system generates a fault prediction report with confidence rating through time series anomaly detection and Bayesian network root cause analysis.

2. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The dynamic resource prediction module includes a heterogeneous resource association analysis unit, a burst load detection unit and a prediction self-correction unit; The heterogeneous resource association analysis unit establishes a nonlinear regression model between GPU memory usage and CUDA core occupancy; The burst load detection unit uses an improved isolation forest algorithm to identify abnormal resource demand spikes; The prediction self-correction unit automatically triggers online fine-tuning of LSTM network parameters when the actual usage deviates from the predicted value by more than a threshold.

3. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The K8s cluster management and resource scheduling module includes a topology-aware scheduling module, and sets an elastic preemption mechanism and an energy-sensitive scheduling strategy; The topology-aware scheduling module optimizes the affinity deployment of multiple containers for AI training tasks based on the connection status between nodes; The elastic preemption mechanism is a dynamic recycling method for low-priority task containers to retain checkpoints and release resources when receiving high-priority requests; The energy consumption-sensitive scheduling strategy dynamically adjusts the node selection weight in combination with the PUE value of the data center.

4. The intelligent AI platform resource management and scheduling system based on K8s according to claim 3 is characterized in that: The execution process of the topology-aware scheduling module is as follows: Automatically identify AllReduce operation mode based on NCCL communication library features; Prioritize assigning tasks of the parameter server architecture to node groups with RDMA network connections; Dynamically adjust the ratio of compute nodes to parameter nodes in mixed precision training scenarios.

5. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The resource isolation and sharing module includes a video memory dynamic reorganization module, a security isolation enhancement layer and a sharing conflict arbitration module; The memory dynamic reorganization module implements cross-container time-sharing multiplexing of GPU memory through virtual address remapping; The security isolation enhancement layer injects an eBPF program to monitor the CUDA API call path when the container is running; The sharing conflict arbitration module dynamically allocates computing time periods based on an auction mechanism when multiple tasks compete for GPU resources.

6. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The computing resource allocation module is based on a dynamic power consumption-performance balance algorithm driven by reinforcement learning. The execution process of automatically adjusting the computing resource allocation strategy when the container is running is as follows: Real-time collection of container dynamic environment data and modeling: Real-time collection of container-level computing power indicators, including GPU utilization, video memory occupancy, and CUDA core activity, to build a multi-dimensional resource status vector; Establishing a power consumption-performance correlation model, and quantitatively calculating the training throughput corresponding to each watt of power consumption through the power consumption-performance correlation model; Predict task load fluctuation trends through time series analysis to form a dynamic decision space; Building a reinforcement learning decision framework: Design a triple model of state space, action space, and reward function. The state space includes cluster resource status, the action space includes computing power allocation strategy, and the reward function includes energy efficiency ratio and task completion rate. Use a deep deterministic policy gradient algorithm to generate computing power allocation plans in the continuous action space; Deploy an online learning mechanism to dynamically update policy network parameters based on scheduling results; Perform real-time strategy execution and feedback: Dynamically adjust container resource limits through the Kubelet extension component to achieve millisecond-level response for GPU computing power allocation; Monitor the actual energy efficiency ratio after the strategy is executed, and mark the results that deviate from the expected value by more than 5% as negative samples; Trigger the resource recovery mechanism to immediately reallocate resources occupied by inefficient allocation strategies.

7. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The security-enhanced image repository includes a blockchain evidence storage and traceability module, a zero-trust vulnerability scanning module, and a factor signature verification module; The blockchain evidence storage and traceability module implements Merkle tree structure encryption on the image layer metadata, generates a unique hash value for each layer and anchors it to the consortium chain, records the hash values ​​of all dependent components from the base image to the final build, and forms an unalterable build history track. When the image is pulled, the on-chain verification process is automatically triggered to compare the consistency of the local image hash with the on-chain evidence. The zero-trust vulnerability scanning module dynamically generates customized vulnerability scanning rules based on the characteristics of the container operating environment (such as kernel version and dependency library list), uses graph neural networks to identify vulnerability propagation paths between image layers, predicts potential attack surfaces, and automatically triggers sandbox environment testing for high-risk images to capture hidden vulnerabilities; The signature verification module automates the generation, rotation, and revocation of signature keys through the HSM hardware security module, embeds K8s cluster fingerprint information when signing the image, ensures that only specific clusters can decrypt and run, and synchronously verifies the image signature and node identity certificate when the container starts, blocking man-in-the-middle attacks.

8. The intelligent AI platform resource management and scheduling system based on K8s according to claim 1 is characterized in that: The inference warning system collects container-level inference latency, GPU memory occupancy, and CUDA core utilization indicators in real time through a multi-dimensional time series data collection and preprocessing unit, constructs a multi-dimensional time series data stream, and performs noise reduction and smoothing on the original data through a sliding window mechanism to eliminate instantaneous interference noise; A dynamic time warping algorithm is used to align timestamps of heterogeneous data sources, ensuring synchronous analysis of multi-dimensional data. A composite anomaly detection model is used for statistical baseline detection, and a Gaussian mixture model is established to characterize the normal fluctuation range of each indicator, marking anomalies that deviate from the 3σ threshold. Deep learning is used for early warning, using a time-domain convolutional network (TCN) to capture long-term pattern anomalies and identify gradual faults that are difficult to detect with traditional methods. The correlation analysis engine calculates the Pearson correlation coefficient between abnormal indicators and screens out abnormal clusters with strong correlations; Dynamic Bayesian root cause inference is established, and a three-layer Bayesian network is constructed, including hardware, framework, and application layer nodes. A conditional probability table is trained based on historical fault data to quantify the causal influence strength between nodes. When an anomaly is detected, the maximum a posteriori probability path is traced back using a variable elimination algorithm to locate the most likely root cause node. Confidence levels are rated and reports are generated. Monte Carlo sampling is used to calculate the confidence interval of the root cause inference results, and the confidence levels are divided into high, medium, and low levels according to the probability values. Generate a structured report, including anomaly timeline diagram, root cause path derivation, and disposal suggestions; The early warning information is pushed to the operation and maintenance platform in real time through the API gateway, triggering automatic repair or manual intervention processes.

Images