Artificial intelligence-oriented cloud computing computing power dynamic allocation method

Abstract

The application relates to the technical field of artificial intelligence, in particular to a cloud computing computing power dynamic allocation method for artificial intelligence. Computing power shrinkage: when the task execution progress is lower than 30% and the computing power node utilization rate is lower than 40%, redundant computing power resources are automatically released, the resources are returned to a resource pool for use by other tasks, and a hibernation operation is performed on the released idle nodes to reduce idle energy consumption; load balancing: when the load difference of computing power nodes in the cluster exceeds 50%, part of the low-priority tasks on the high-load nodes are migrated to the low-load nodes to realize cluster load balancing and avoid energy consumption soaring caused by single node overload; computing power adaptation precision: by constructing an AI task multi-dimensional feature model and a heterogeneous computing power resource pooling model, the task computing power adaptation tendency and resource characteristics are deeply mined to realize precise matching of AI tasks and heterogeneous computing power, and the computing power supply-demand mismatch rate is reduced by more than 40%; operation guarantee stabilization: adaptive migration and resource recycling guarantee continuous execution of AI tasks.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for dynamic allocation of cloud computing power for artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI applications such as large language models, computer vision, and intelligent recommendation are exhibiting heterogeneous, dynamic, and exponentially growing demands for cloud computing power: training tasks are computationally intensive, requiring massive GPU / TPU clusters for parallel computation; inference tasks emphasize low latency and high throughput, requiring flexible adaptation of heterogeneous computing power such as CPU / GPU / FPGA; and edge AI tasks require localized scheduling of computing power and low-energy operation. Meanwhile, the scale of heterogeneous computing power on cloud computing platforms continues to expand, leading to rising energy costs. Existing allocation methods lack targeted energy-saving management, further exacerbating resource waste and cost burdens.

[0003] Existing cloud computing power allocation methods have the following core flaws:

[0004] Static allocation is the main method, resulting in poor adaptability: It often adopts fixed quota allocation or threshold-triggered elastic scaling, which cannot match the dynamic fluctuations in computing power requirements of AI tasks, and is prone to resource idleness or overload.

[0005] Insufficient heterogeneous computing power scheduling capability: The heterogeneous computing power such as CPU, GPU, FPGA, and ASIC is not uniformly modeled and coordinated for scheduling, resulting in problems such as computing power fragmentation and low utilization of dedicated computing power, making it difficult to support complex AI tasks such as distributed training of large models and real-time inference.

[0006] Insufficient task feature mining: Resource allocation is based solely on task computation volume without in-depth mining of multi-dimensional features of AI tasks such as model size, data dimensions, accuracy requirements, and latency constraints. In particular, it fails to accurately match the computing power adaptation tendency of the task, resulting in a mismatch between computing power supply and demand.

[0007] Scheduling decisions are lagging and global optimality is poor: Traditional scheduling algorithms rely on static rules or single-point-in-time load snapshots, lacking prediction of computing load trends and global optimization, making them prone to getting trapped in local optima and difficult to balance resource utilization, task execution efficiency and energy consumption costs.

[0008] Lack of energy conservation management: No dedicated energy conservation strategies were designed for the computing power requirements of AI tasks. Heterogeneous computing nodes were in a state of full load or inefficient operation for a long time, resulting in serious energy waste and failing to achieve the dual goal of "high-efficiency computing power + low-consumption operation".

[0009] In summary, existing technologies cannot meet the core requirements of AI applications for precise matching, dynamic scheduling, heterogeneous collaboration, and high efficiency and low consumption of cloud computing power. There is a need for a dynamic allocation method of cloud computing power for artificial intelligence, to replace traditional technical terminology, supplement energy-saving management technology, solve the above-mentioned technical defects, and support the large-scale and low-cost application of AI technology. Summary of the Invention

[0010] This invention aims to overcome the shortcomings of existing technologies and provide a dynamic allocation method for cloud computing power for artificial intelligence. It replaces the technical descriptions related to "inference tasks" and "multi-dimensional features of resource preferences" and adds a full-process computing power energy-saving management and control technology. This enables accurate modeling of computing power for different types of AI tasks, collaborative scheduling of heterogeneous resources, dynamic load adaptive optimization and energy-saving operation, improves the utilization rate of computing power resources of cloud computing platforms and the execution efficiency of AI tasks, and reduces computing power scheduling costs, task latency and energy consumption.

[0011] To achieve the above objectives, this invention provides a method for dynamic allocation of cloud computing power for artificial intelligence, comprising the following steps:

[0012] Step 1: Multidimensional Feature Extraction and Modeling for AI Tasks

[0013] Receive AI task requests submitted by users, extract core task features, and construct a multi-dimensional feature vector. The feature dimensions include:

[0014] Task types: training tasks, inference tasks, edge AI tasks;

[0015] Model parameters: model size, network structure, accuracy requirements;

[0016] Data characteristics: data size, data dimensions, batch size, number of iterations;

[0017] Constraints: latency threshold, throughput requirement, energy consumption limit, priority level;

[0018] Computing power adaptation tendency: computing power type adaptation tendency, video memory requirements, memory requirements, network bandwidth requirements, quantitatively characterizing the adaptation priority of tasks to different types of computing power.

[0019] After quantizing the above multidimensional features, construct the AI ​​task feature matrix T=[t1,t2,...,t], where tᵢ is the feature vector of the i-th AI task and n is the number of tasks.

[0020] Step 2: Modeling and State Awareness of Heterogeneous Computing Resources Pooling

[0021] Unified pooling management of heterogeneous computing resources on cloud computing platforms is implemented to build a computing resource pool. Resource types include: general-purpose CPU nodes, GPU nodes, FPGA nodes, ASIC-specific acceleration nodes, and edge computing nodes.

[0022] Real-time collection of computing resource pool status data is used to construct a resource status matrix R=[r1,r2,...,r], where r is the status vector of the j-th computing node, m is the number of nodes, and the status vector includes:

[0023] Hardware parameters: number of CPU cores, GPU memory size, number of FPGA logic units, memory capacity, storage capacity;

[0024] Real-time load: CPU utilization, GPU utilization, memory usage, network bandwidth utilization, number of tasks running;

[0025] Performance metrics: floating-point arithmetic capability, data throughput, average latency, and power consumption;

[0026] Topology information: node geographical location, network link bandwidth, cluster affiliation;

[0027] Energy saving potential parameters: node idle energy consumption, energy consumption coefficient under different loads, and energy saving adjustment threshold.

[0028] At the same time, a computing resource-task adaptation mapping library is built to store historical adaptation data and energy consumption data of different types of AI tasks and heterogeneous computing nodes, providing a reference for subsequent scheduling decisions and energy-saving management.

[0029] Step 3: Dynamic computing power allocation decision based on deep reinforcement learning

[0030] A deep reinforcement learning scheduling model is constructed, with the goal of minimizing the average execution latency of AI tasks, maximizing the utilization of computing resources, and minimizing energy consumption costs as the multi-objective optimization function. Energy-saving constraints are incorporated to output the optimal computing power allocation strategy.

[0031] DRL Model Construction

[0032] State space S: The input consists of the AI ​​task feature matrix T, the computing power resource state matrix R, and the energy-saving potential parameters of each node, which are fused to form a global state vector;

[0033] Action Space A: The output is computing power allocation actions, including: task-node matching, computing power quota allocation, task scheduling order, resource pre-allocation strategy, and energy-saving adjustment actions;

[0034] Reward function R: Design a multi-dimensional reward mechanism, adding energy-saving related rewards and penalties, including:

[0035] Positive rewards: Task execution latency is below the threshold (reward +10), computing resource utilization is above 80% (reward +5), high-priority tasks are executed first (reward +8), node energy consumption is below the energy saving threshold (reward +7), idle resources successfully go into sleep (reward +6).

[0036] Negative penalties: Task execution latency exceeds the limit (penalty -15), computing power resource overload (penalty -10), severe resource fragmentation (penalty -5), node energy consumption exceeds the limit (penalty -8), invalid wake-up of dormant nodes (penalty -4).

[0037] Network structure: A deep neural network (DNN) is used as the policy network, a convolutional neural network (CNN) is used to extract task and resource features, and a near-end policy optimization (PPO) algorithm is used in the reinforcement learning algorithm to improve the model training stability and convergence speed. A new energy-saving feature extraction branch is added to optimize the accuracy of energy-saving decision-making.

[0038] Scheduling decision execution

[0039] The AI ​​task feature matrix and computing resource state matrix are input into the trained DRL model, and the model outputs the optimal computing power allocation strategy, including:

[0040] Task-node matching: Combining task computing power adaptation tendency, AI tasks are assigned to heterogeneous computing power nodes with the highest adaptability and the best energy saving potential (such as large model training tasks being assigned to multi-GPU clusters, and real-time inference tasks being assigned to low-latency CPU / GPU nodes with low energy consumption).

[0041] Precise allocation of computing power quotas: Dynamically allocate computing power resources based on task characteristics (e.g., adjust GPU memory quotas for training tasks according to iteration stages, and adjust the number of CPU cores for inference tasks according to request concurrency) to avoid energy waste caused by resource redundancy;

[0042] Task priority scheduling: High-priority tasks occupy high-quality computing resources first, while low-priority tasks can be scheduled or processed in batches when resources are idle, reducing the consumption of high-energy-consuming resources by low-priority tasks.

[0043] Resource pre-allocation: Based on historical data, predict future peak computing power demand and pre-allocate redundant resources in advance to avoid latency exceeding limits due to sudden requests, while controlling the scale of pre-allocated resources to reduce idle energy consumption;

[0044] Initial energy-saving decision: Based on node load and energy-saving potential, idle nodes are put into hibernation mode, and low-load nodes are frequency-reduced to initially reduce global energy consumption.

[0045] Step 4: Real-time load feedback and elastic optimization of computing power

[0046] During the execution of AI tasks, the system monitors the task's running status, changes in computing node load, and energy consumption data in real time, constructing a closed-loop feedback mechanism to achieve flexible optimization and dynamic adjustment of computing resources for energy conservation.

[0047] Real-time status monitoring

[0048] The following data is collected in real time through a monitoring agent deployed on the computing power nodes:

[0049] Task running status: task execution progress, remaining computing power requirement, current latency, error log;

[0050] Computing node load: CPU / GPU utilization, memory usage, network bandwidth utilization, and energy consumption fluctuations;

[0051] Cluster global status: average resource utilization, task queue length, latency distribution, and degree of resource fragmentation;

[0052] Energy-saving operation status: deviation between actual node energy consumption and energy-saving threshold, wake-up requirements of dormant nodes, and impact on the performance of frequency-reduced nodes.

[0053] Flexible tuning and energy-saving synergy strategy

[0054] Based on real-time monitoring data, the following coordinated actions for flexible tuning and energy saving are triggered:

[0055] Computing power expansion: When the task latency continues to exceed the threshold and the computing power node utilization exceeds 90%, the computing power quota will be automatically increased (such as adding GPU cards or expanding the number of CPU cores) or idle computing power nodes will be scheduled to join the task execution. When waking up dormant nodes, nodes with high energy saving potential will be given priority.

[0056] Computing power reduction: When the task execution progress is less than 30% and the computing power node utilization is less than 40%, redundant computing power resources are automatically released and returned to the resource pool for use by other tasks. At the same time, the idle nodes after release are put into hibernation to reduce idle energy consumption.

[0057] Load balancing: When the load difference between computing nodes in the cluster exceeds 50%, some low-priority tasks on high-load nodes are migrated to low-load nodes to achieve cluster load balancing and avoid energy consumption spikes caused by overload of a single node.

[0058] Dynamic energy-saving adjustment: Under the premise of meeting task delay constraints, dynamic frequency reduction is performed on low-load nodes (utilization rate 20%-40%) to reduce energy consumption; nodes that are idle for more than 30 minutes are automatically put into hibernation; for nodes with excessive energy consumption, task allocation is adjusted to reduce node load to within the energy-saving threshold.

[0059] Accuracy-adaptive energy saving: For inference tasks, while meeting accuracy constraints, the model accuracy is dynamically adjusted (e.g., from FP16 to INT8) to reduce computing power requirements and thus reduce energy consumption.

[0060] Step 5: Adaptive Migration and Resource Reclamation of Abnormal Tasks

[0061] To address abnormal situations during AI task execution (such as task crashes, computing node failures, and insufficient computing resources), the system enables adaptive migration and resource recovery of abnormal tasks, ensuring continuous task execution and efficient resource utilization while avoiding energy waste caused by abnormal states.

[0062] Anomaly detection

[0063] Real-time detection of abnormal states through monitoring data, including:

[0064] Task errors: Task execution is interrupted, errors occur, progress stalls, and latency continues to spike;

[0065] Node anomalies: computing node crashes, hardware failures, network interruptions, abnormal loads, and abnormal spikes in energy consumption;

[0066] Resource anomalies: exhaustion of computing resources, memory overflow, insufficient video memory.

[0067] Adaptive migration and resource recycling

[0068] Abnormal Task Migration: When an abnormal task or node failure is detected, the task execution context (such as model parameters, training progress, and data cache) is immediately saved, and the task is migrated to a normal, compatible computing power node in the resource pool (with priority given to nodes with high energy-saving potential) to resume task execution and ensure uninterrupted task operation.

[0069] Forced resource reclamation: For abnormal tasks that cannot be recovered, the computing resources they occupy are immediately released and returned to the resource pool; for computing resources that have been idle for a long time, they are automatically reclaimed and redistributed to avoid resource waste and ineffective energy consumption.

[0070] Fault node isolation: Isolate the faulty computing node from the resource pool, trigger an alarm and start the maintenance process to prevent the fault from spreading and affecting other tasks, and at the same time stop the invalid energy supply of the faulty node.

[0071] Step 6: Full-process energy-saving management and optimization iteration

[0072] A new end-to-end energy-saving management module has been added, covering the entire process of computing power allocation, task execution, and resource recycling, to achieve continuous optimization of energy-saving strategies. Specifically, this includes:

[0073] Energy-saving baseline establishment: Based on the computing power requirements and historical energy consumption data of different types of AI tasks, establish energy-saving baselines for each task and node (such as inferring task energy consumption thresholds and GPU node idle energy consumption standards) as a reference for energy-saving management.

[0074] Real-time energy consumption statistics and analysis: Real-time statistics of energy consumption data for each computing node and each AI task, comparison with energy-saving baseline, analysis of the causes of abnormal energy consumption (such as resource redundancy, uneven load, node failure), and generation of energy consumption analysis report;

[0075] Iterative optimization of energy-saving strategies: Feed the energy consumption analysis results back to the DRL scheduling model to continuously optimize the energy-saving weights, node matching strategies and dynamic energy-saving adjustment parameters in the reward function, improve the energy-saving effect, and ensure that the task execution efficiency is not affected.

[0076] Energy saving assessment and feedback: Regularly assess the energy saving effect of the cloud computing platform, including the balance between energy consumption reduction rate, resource utilization rate and energy saving effect, and generate assessment reports to support the optimization of subsequent energy saving strategies.

[0077] The beneficial effects of this invention are:

[0078] Accurate computing power matching: By constructing a multi-dimensional feature model of AI tasks and a heterogeneous computing power resource pooling model, we can deeply mine the computing power matching tendency and resource characteristics of tasks, achieve accurate matching between AI tasks and heterogeneous computing power, and reduce the computing power supply and demand mismatch rate by more than 40%.

[0079] Highly efficient resource utilization: Based on dynamic scheduling decisions using deep reinforcement learning, combined with real-time load feedback and elastic optimization, dynamic allocation and efficient utilization of computing resources are achieved, increasing resource utilization from the traditional 60%-70% to over 85%.

[0080] Low latency in task execution: Through priority scheduling, resource pre-allocation, load balancing and anomaly migration, the latency of AI task execution is effectively reduced, with training task latency reduced by more than 25% and inference task latency reduced by more than 30%.

[0081] Heterogeneous scheduling coordination: Enables unified pooling management and coordinated scheduling of heterogeneous computing resources such as CPU, GPU, FPGA, and ASIC, solves the problem of computing power fragmentation, and supports complex scenarios such as distributed training of large models and edge AI computing;

[0082] Intelligent scheduling process: It adopts a deep reinforcement learning model to learn scheduling rules autonomously without human intervention, automatically adapts to the dynamic changes in the computing power requirements of AI tasks, and the scheduling decision has strong global optimality, avoiding the trap of local optima;

[0083] Operational stability is ensured: Through anomaly detection, adaptive migration and resource reclamation, the continuous execution of AI tasks is guaranteed, the task interruption rate is reduced to below 0.5%, and the operational stability is significantly improved;

[0084] Significant energy-saving effect: The new full-process energy-saving management and control technology, combined with strategies such as dynamic frequency reduction, node hibernation, and precision adaptive adjustment, realizes the synergy between computing power allocation and energy-saving management and control, reducing the overall energy consumption cost by more than 40%, solving the problem of serious energy waste in existing technologies, and achieving the goal of "high efficiency, low consumption, and stability" computing power scheduling.

[0085] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0086] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0088] The present invention is illustrated below with specific embodiments, but these are not intended to limit the scope of the invention. Figure 1 As shown, a dynamic allocation method for cloud computing power for artificial intelligence is proposed.

[0089] Example 1: Computational Power Allocation and Energy Saving Management for Large Language Model Training Tasks

[0090] Task characteristics: Training task, GPT-3 class model (175 billion parameters), FP16 accuracy, 100,000 training iterations, batch size 32, latency constraint ≤24 hours, high priority, computing power adaptation is preferred to multi-GPU cluster;

[0091] Resource pool status: It includes 10 GPU cluster nodes (each node has 8×NVIDIA H100 GPUs and 80GB of video memory), 20 CPU nodes, and 5 FPGA nodes. The GPU cluster utilization is 60%, the CPU node utilization is 40%, the energy-saving baseline for each node has been established, and the idle energy consumption of the GPU nodes is 30% of the full load.

[0092] Scheduling decision: The DRL model is matched to GPU cluster nodes, with 8×H100 GPUs / node allocated, for a total of 5 nodes. 20% redundant GPU resources are pre-allocated to cope with iteration peaks, and the highest priority is given to scheduling. The 5 idle GPU nodes are put into sleep mode to reduce idle energy consumption.

[0093] Flexible tuning and energy-saving adjustment: During the mid-training stage, when GPU utilization reaches 95%, two GPU nodes are automatically added; during the late-training stage, when utilization drops to 50%, three nodes are reduced to release resources and put the system to sleep; the remaining nodes are dynamically downclocked based on the load, reducing energy consumption by 30%.

[0094] Results: Model training was completed in 22 hours, with 92% utilization of computing resources, no task interruptions, and energy costs reduced by 38%, which is better than the preset energy-saving baseline.

[0095] Example 2: Computational Power Allocation and Energy Saving Management for Real-Time Image Recognition and Inference Tasks

[0096] Task characteristics: Inference task, CNN model (ResNet-50), INT8 accuracy (after adaptive power saving adjustment), peak concurrent requests 1000 QPS, latency constraint ≤100ms, medium priority, computing power adaptation tends to low latency GPU nodes;

[0097] Resource pool status: Contains 5 GPU inference nodes (NVIDIA A10 GPU, 24GB VRAM) and 10 CPU nodes (Intel Xeon 8375C). GPU utilization is 50%, CPU utilization is 60%, and the GPU node power-saving threshold is 60% of full load power consumption.

[0098] Scheduling decisions: The DRL model is matched to the GPU inference node, 4×A10 GPUs are allocated, the GPU memory quota is dynamically adjusted according to the concurrency, and a low-latency scheduling strategy is adopted; the CPU nodes with low load are downclocked to reduce energy consumption.

[0099] Elastic tuning and energy saving: When GPU utilization reaches 90% during peak periods, one GPU node is automatically added; when utilization drops to 30% during off-peak periods, the node is scaled down and put into sleep mode, and requests are distributed in a load-balanced manner; at the same time, the model accuracy is dynamically adjusted to further reduce energy consumption while meeting the recognition accuracy requirements.

[0100] Results: The average latency was 85ms, the computing resource utilization rate was 88%, the request throughput met the target, there were no cases of latency exceeding the target, and the energy consumption cost was reduced by 42%, achieving the dual goals of high efficiency and energy saving.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic allocation of cloud computing power for artificial intelligence, characterized in that, Includes the following steps: Step 1, AI task multidimensional feature extraction and modeling: Receive AI task requests, extract multidimensional features such as task type, model parameters, data features, constraints, and computing power adaptation tendency, and quantitatively construct the AI ​​task feature matrix; The task types include training tasks, inference tasks, and edge AI tasks; The computing power adaptation tendency characterizes the priority of task adaptation to different types of computing power; Step 2, Heterogeneous computing power resource pooling modeling and status awareness: The heterogeneous resources of the cloud computing platform, including CPU, GPU, FPGA, ASIC, and edge computing nodes, are pooled in a unified manner. Hardware parameters, real-time load, performance indicators, topology information, and energy-saving potential parameters are collected in real time to construct a resource status matrix. Step 3, Dynamic computing power allocation decision based on deep reinforcement learning: Construct a deep reinforcement learning scheduling model with multiple objectives of minimizing task latency, maximizing resource utilization, and minimizing energy consumption. Input the task feature matrix, resource state matrix, and energy-saving potential parameters, and output the optimal strategies for task-node matching, computing power quota allocation, priority scheduling, resource pre-allocation, and energy-saving adjustment. Step 4, Real-time load feedback and elastic optimization of computing power: Real-time monitoring of task running status, computing power node load and energy consumption data, triggering elastic actions such as computing power expansion, reduction, load balancing and dynamic energy saving adjustment to achieve dynamic resource adjustment and energy saving coordination. Step 5, Adaptive Migration and Resource Reclamation for Abnormal Tasks: Detect abnormal states such as task crashes, node failures, insufficient resources, and abnormal spikes in energy consumption, save the task context, migrate to normal computing power nodes, reclaim redundant resources, isolate faulty nodes, and stop their ineffective energy consumption. Step 6, Full-process energy-saving management and optimization iteration: Establish energy-saving baseline, collect and analyze energy consumption data in real time, feed the analysis results back to the deep reinforcement learning scheduling model, iteratively optimize energy-saving strategies, and regularly evaluate the energy-saving effect.

2. The method according to claim 1, characterized in that, The inference task described in step 1 includes real-time image recognition, text question answering, and batch data processing; The model parameters include model size, network structure, and accuracy requirements; the constraints include latency threshold, throughput requirements, energy consumption limits, and priority levels.

3. The method according to claim 1, characterized in that, The deep reinforcement learning scheduling model described in step 3 adopts the near-end policy optimization PPO algorithm. The state space is a fusion vector of task, resource and energy-saving features, the action space is a set of computing power allocation and energy-saving adjustment actions, and the reward function includes positive rewards for task delay, resource utilization, priority and energy consumption and negative penalties for resource overload, fragmentation and energy consumption exceeding the standard.

4. The method according to claim 1, characterized in that, The triggering conditions for the coordinated elastic tuning and energy saving mentioned in step 4 include: task latency continuously exceeding the threshold and computing node utilization exceeding 90% to trigger capacity expansion; When the task progress is below 30% and the node utilization is below 40%, scaling down and node hibernation will be triggered. A load balancing event is triggered when the cluster load difference exceeds 50%. Dynamic frequency reduction is triggered when node utilization reaches 20%-40%. A node will trigger a hibernation event if it remains idle for more than 30 minutes.

5. The method according to claim 1, characterized in that, The anomaly detection described in step 5 includes task execution interruption, progress stagnation, node crash, hardware failure, resource exhaustion, and abnormal surge in energy consumption; The adaptive migration needs to save model parameters, training progress, data cache execution context to ensure uninterrupted task recovery, and prioritize computing nodes with high energy-saving potential during migration.

6. The method according to claim 1, characterized in that, The energy-saving adjustment mentioned in step 6 includes dynamic frequency reduction of nodes, hibernation of idle nodes, and adaptive adjustment of the accuracy of the inference task model; The energy-saving baseline is established based on task type, node model, and historical energy consumption data, and is used to evaluate the effectiveness of energy consumption optimization.

Images