Intelligent computing center resource dynamic planning method based on artificial intelligence prediction

Abstract

The invention belongs to the field of computing resource planning, relates to an artificial intelligence prediction technology, and aims to solve the problem that a load prediction result cannot be dynamically optimized according to a task execution state in the prior art, in particular to an intelligent computing center resource dynamic planning method based on artificial intelligence prediction. Comprising the following steps: carrying out periodic prediction analysis on the load of a computing center: generating a prediction period, dividing the prediction period into a plurality of prediction time periods, and obtaining task queue historical data, resource utilization rate historical data and metadata at the beginning time of the prediction time periods; the task queue historical data, the resource utilization rate historical data and the metadata are aligned and aggregated according to the unified time frequency; according to the method, intelligent dynamic planning of computing resources is realized, the system can quickly respond to load change through periodic load prediction and resource demand interval division, and the problem of resource fragmentation caused by fixed-scale cluster configuration is avoided.

Description

Technical Field

[0001] This invention belongs to the field of computing resource planning and relates to artificial intelligence prediction technology, specifically a dynamic planning method for intelligent computing center resources based on artificial intelligence prediction. Background Technology

[0002] The core of the intelligent computing center resource dynamic planning method lies in leveraging artificial intelligence technology to achieve a fundamental shift from "passive response" to "proactive anticipation." It achieves efficient, green, and automated operation of computing resources by accurately predicting future workloads and proactively planning and scheduling resources intelligently.

[0003] The invention patent with publication number CN113703952B discloses a resource allocation method based on supercomputer queue resource scheduling. This allocation method optimizes the configuration of computing resources, improves efficiency, and can maintain a robust resource queue for resource calls in emergency situations. However, this allocation method can only allocate tasks based on the computing load prediction results and the saturation status of individual computing nodes. It cannot perform a holistic analysis of the task execution status of the computing cluster composed of computing nodes. This results in unresolved issues such as task blocking caused by memory allocation conflicts in GPU clusters and task backlog caused by uneven computing power among computing nodes in heterogeneous clusters (CPU+GPU). In addition, the existing technology cannot dynamically optimize the load prediction results based on the task execution status, resulting in the inability to guarantee the timely completion rate of computing tasks.

[0004] To address the aforementioned technical problems, this application proposes a solution. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction, which solves the problem that existing technologies cannot dynamically optimize load prediction results according to task execution status; The technical problem to be solved by this invention is: how to provide a dynamic resource planning method for intelligent computing centers based on artificial intelligence prediction that can dynamically optimize load prediction results according to task execution status.

[0006] The objective of this invention can be achieved through the following technical solutions: The method for dynamic resource planning in intelligent computing centers based on artificial intelligence prediction includes the following steps: Step S1: Perform periodic predictive analysis on the load of the computing center: Generate a prediction period and divide the prediction period into several prediction periods. At the beginning of each prediction period, acquire historical data of the task queue, historical data of resource utilization, and metadata. Align and aggregate the historical data of the task queue, historical data of resource utilization, and metadata according to a uniform time frequency. Input the historical data of the task queue, historical data of resource utilization, and metadata into a recurrent neural network for load analysis to obtain the peak resource demand and allocation quantity for the prediction period. Step S2: Assign tasks to the computing nodes in the computing center: randomly select reserved objects and form a reserved processing cluster from the reserved objects; Step S3: Analyze the rationality of task allocation in the computing center: At the end of the prediction period, obtain the distribution coefficient of the reserved processing cluster; at the end of the prediction period, count the peak resource demand for all prediction periods; the maximum and minimum values ​​of the peak resource demand constitute the resource demand range; divide the resource demand range into several resource demand intervals; mark the configuration clusters in the resource demand intervals; when the peak resource demand is obtained in the prediction period of the next prediction period, retrieve the configuration clusters in the resource demand interval corresponding to the peak resource demand as the reserved processing clusters. Step S4: Analyze the task execution status of the computing center.

[0007] Further, in step S1, the historical task queue data is high-frequency time-series data collected from the scheduler. The task queue data includes timestamps, the number of submitted tasks, the number of tasks in the queue, and the number of tasks running. The historical resource utilization data is collected from the monitoring system. The historical resource utilization data includes total CPU utilization, memory usage, and GPU utilization. The metadata includes average task duration and requested CPU / GPU / memory.

[0008] Furthermore, in step S2, the specific process of task allocation for the computing nodes of the computing center includes: marking the computing nodes of the computing center as allocation objects, randomly selecting allocation objects as reserved objects, the number of reserved objects being equal to the number of allocation objects, forming a reserved processing cluster from the reserved objects, and evenly distributing the computing tasks received during the prediction period to the reserved objects in the reserved processing cluster for processing according to the amount of computing.

[0009] Further, in step S3, the process of obtaining the uniformity coefficient includes: counting the total time for each reserved object to perform the calculation task and marking it as the processing value of the reserved object, and calculating the variance of the processing values ​​of all reserved objects within the prediction period to obtain the uniformity coefficient.

[0010] Furthermore, in step S3, the specific marking process for the configuration clusters of the resource demand interval includes: marking the reserved processing clusters of the prediction period whose resource demand peak is within the resource demand interval as the matching clusters of the resource demand interval, and marking the matching cluster with the smallest distribution coefficient as the configuration clusters of the resource demand interval.

[0011] Furthermore, in step S4, the specific process of analyzing the task execution status of the computing center includes: at the end of the prediction period, summing and averaging the processing values ​​of all reserved objects in the configuration cluster corresponding to the resource demand interval to obtain the execution coefficient of the resource demand interval, and judging whether the task execution status of the configuration cluster in the resource demand interval meets the requirements by using the execution coefficient. If the requirements are not met, the execution optimization analysis is performed on the extended pre-determined interval.

[0012] Furthermore, the specific process for determining whether the task execution status of the configuration cluster in the resource demand range meets the requirements includes: comparing the execution coefficient with a preset execution threshold; if the execution coefficient is less than the execution threshold, the task execution status of the configuration cluster in the resource demand range is determined to meet the requirements; if the execution coefficient is greater than or equal to the execution threshold, the task execution status of the configuration cluster in the resource demand range is determined to not meet the requirements, and the corresponding resource demand range is marked as an extended range.

[0013] Furthermore, the specific process of performing optimization analysis on the extended forecast interval includes: marking the number of times the allocated object is included in the reserved processing cluster within the current forecast period as the inclusion value; when the peak resource demand during the forecast period in the next forecast period is within the extended forecast interval, the allocation update value FPg is obtained through the formula FPg=[t1×FP], where t1 is the proportional coefficient and 1.15≤t1≤1.25, and FP is the value of the allocation quantity; the allocated object with the largest inclusion value other than the reserved object in the configuration cluster corresponding to the extended forecast interval is added to the configuration cluster until the number of reserved objects in the configuration cluster reaches the allocation update value FPg.

[0014] The present invention has the following beneficial effects: 1. This application realizes intelligent dynamic planning of computing resources. Through periodic load forecasting and resource demand interval division, the system can quickly respond to load changes, avoiding the resource fragmentation problem caused by fixed-size cluster configuration. The introduction of the distribution coefficient effectively improves the balanced distribution of tasks in heterogeneous clusters, reducing task blocking and backlog caused by GPU memory conflicts or CPU / GPU computing power imbalance. Through cluster analysis and feedback mechanism of historical execution data, the load forecasting model can continuously optimize itself and improve the prediction accuracy. This closed-loop optimization system significantly improves the resource utilization of the computing center, reduces task execution latency, and reduces energy consumption, providing strong support for the efficient and green operation of the computing center. 2. This application achieves accurate prediction of computing center load. By collecting and integrating multi-dimensional historical data and combining it with a recurrent neural network model, it can accurately predict the peak resource demand and allocation quantity in the future. This prediction method considers multiple factors such as task queues, resource utilization, and task characteristics, which improves the accuracy and comprehensiveness of the prediction. At the same time, by defining reasonable prediction targets and time windows, and by encapsulating the model into an API and triggering it through scheduled tasks, the prediction process is automated and real-time, providing reliable data support for subsequent dynamic resource planning. 3. This application achieves a balanced distribution of computing tasks within the reserved processing cluster. This method avoids overload of individual computing nodes and improves overall resource utilization efficiency. At the same time, since the reserved objects are randomly selected, this method also increases the flexibility and robustness of the system, enabling it to adapt to different types and scales of computing tasks. Furthermore, the size of the reserved processing cluster is dynamically adjusted based on prediction results, which ensures that resource allocation matches actual needs and avoids problems of resource waste or insufficiency. 4. This application realizes dynamic analysis and optimization of the task execution status of the computing center. By introducing the execution coefficient as a quantitative indicator, the actual load of each resource demand interval can be objectively evaluated. Furthermore, based on the comparison between the execution coefficient and the preset threshold, the overloaded interval can be identified in a timely manner and marked as the expansion interval. This mechanism enables the system to proactively discover potential performance bottlenecks and provide a basis for subsequent resource adjustments, thereby improving the overall operating efficiency and resource utilization of the computing center. At the same time, since the analysis process is carried out at the end of the prediction period, it will not interfere with the execution of real-time tasks, ensuring the stability of the system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Traditional dynamic resource planning methods focus solely on the resource saturation of individual computing nodes, lacking comprehensive analysis of the overall task execution status of the computing cluster. Task blocking issues caused by GPU memory allocation conflicts in heterogeneous computing clusters cannot be effectively identified, and task backlog due to uneven CPU and GPU computing power continues to worsen. Because a feedback mechanism between prediction results and cluster execution performance is not established, the load prediction model cannot dynamically adjust based on historical task execution data, resulting in a systematic deviation between resource allocation strategies and actual needs.

[0019] For example, when running deep learning inference tasks in an intelligent computing center using a CPU+GPU heterogeneous architecture, the monitoring system collects data on CPU utilization, GPU memory usage, and task queue length for each node at a frequency of minutes. The scheduler predicts the peak computing power demand for the next 30 minutes based on an LSTM model and randomly selects nodes that meet the resource thresholds to form a temporary processing cluster. When a sudden image recognition task triggers GPU memory contention, no memory conflict early warning mechanism is established within the cluster. Some nodes enter a blocked state due to memory exhaustion, while CPU nodes experience resource gaps due to idle computing power. Furthermore, because no cluster analysis is performed on the distribution characteristics of historical task execution durations, the predicted peak resource demand always triggers a fixed-size cluster configuration, failing to adapt to dynamically changing load characteristics.

[0020] If the above issues are not addressed, resource fragmentation in heterogeneous computing clusters will cause overall resource utilization to drop beyond design thresholds, and end-to-end latency for critical tasks will exceed service level agreement constraints. Task retry mechanisms triggered by memory allocation conflicts will significantly increase system energy consumption, while localized hotspots caused by uneven computing power may trigger cascading node failures. A lack of dynamically optimized load forecasting models will continuously amplify resource allocation errors, ultimately leading to an exponential increase in computing center operating costs and task failure rates.

[0021] like Figure 1 As shown, the intelligent computing center resource dynamic planning method based on artificial intelligence prediction includes the following steps: Step S1: Perform periodic predictive analysis on the load of the computing center: Generate a prediction period and divide it into several prediction periods. At the beginning of each prediction period, acquire historical task queue data, historical resource utilization data, and metadata. The historical task queue data is obtained from the scheduler (e.g., Slurm). In Kubernetes, high-frequency (e.g., every minute) time-series data is collected. Task queue data includes timestamps, the number of submitted tasks, the number of queued tasks, and the number of running tasks. Historical resource utilization data is collected from monitoring systems (e.g., Prometheus), including total CPU utilization, memory usage, and GPU utilization. Metadata includes average task duration and requested CPU / GPU / memory. The historical task queue data, historical resource utilization data, and metadata are aligned and aggregated at a uniform time frequency (e.g., every 5 minutes). The historical task queue data, historical resource utilization data, and metadata are input into a recurrent neural network for load analysis to obtain the peak resource demand and allocation quantity for the predicted period. The prediction objective is defined as follows: Prediction range: the next L1 minutes; Input window: using data from the past 4 × L1 minutes as input; Output: a result containing the peak resource demand and allocation quantity for the next L1 minutes. The trained model is encapsulated into an API, and a scheduled task (e.g., Airflow) triggers the prediction (at the start time of the prediction period). The model API is called to obtain the prediction result for the next L1 minutes; L1 minutes is the duration of the prediction period. Historical data for the task queue is collected at a second-level frequency by the scheduler, historical data for resource utilization is collected at a minute-level frequency by the monitoring system, and metadata is collected from task submission logs triggered by task completion events. A unified time frequency is achieved by downsampling the second-level data to the minute-level using data interpolation methods to align it with the minute-level data. Simultaneously, event-triggered metadata is aggregated and statistically analyzed using minute-level time windows. The input window for the prediction target is set to a data volume four times the past prediction duration; for example, when L1 is 15 minutes, the input window contains 60 minutes of historical data. Scheduled tasks use the Airflow scheduler to trigger model API calls at the start of the prediction period, and the model output is directly mapped to the future L1-minute timeline.

[0022] Specifically, the scheduler collects the task queue status every 30 seconds, the monitoring system collects CPU, memory, and GPU utilization data every minute, and metadata records the resource request parameters and execution duration when a task is completed. By taking the last value of the task queue data by minute and summing and averaging the metadata by minute, the three types of data are aligned on the minute timestamp. The input window length and prediction range form a 4:1 ratio to ensure that the model can capture the characteristics of periodic load fluctuations. The trained recurrent neural network model is encapsulated as a RESTful API. The Airflow scheduler sends an HTTP request to the API endpoint at the start of the prediction period. Based on the aligned input data, the model generates the predicted peak resource demand value for the next L1 minute and the corresponding number of compute nodes allocated. This allocation number is directly used for the construction of the reserved processing cluster in subsequent steps. For example, when L1 is 15 minutes and the predicted allocation number is 8, the system will dynamically allocate 8 compute nodes to process the task queue in the next 15 minutes.

[0023] Step S2: Assign tasks to the computing nodes in the computing center: Mark the computing nodes in the computing center as assignment objects, randomly select assignment objects as reserved objects, the number of reserved objects is equal to the number of assignment objects, and the reserved objects form a reserved processing cluster. The computing tasks received during the prediction period are evenly distributed to the reserved objects in the reserved processing cluster for processing according to the amount of computing. In this system, computing nodes are uniformly marked as allocation objects to form a resource pool. A random selection mechanism avoids the risk of load concentration caused by fixed node combinations. The number of reserved objects strictly corresponds to the allocation quantity output by the prediction model. For example, when the allocation quantity is FP, FP nodes are randomly selected from the resource pool to form a cluster. Computational tasks are divided into equal-quantity tasks according to their computational complexity and evenly distributed to the nodes within the cluster using a round-robin scheduling algorithm.

[0024] Specifically, at the start of the prediction period, the size of the reserved processing cluster is determined based on the allocation quantity output by the prediction model. A corresponding number of nodes are randomly selected from the set of computing nodes marked as allocation targets to form a dynamic cluster. Computational tasks in the task queue are split into multiple subtasks, with the number of subtasks being an integer multiple of the number of cluster nodes. For example, when the cluster contains 10 nodes, each task is split into 10 subtasks. Subtasks are mapped to different nodes using a hash algorithm to ensure balanced processing load across nodes. This process combines dynamic random selection with uniform allocation to eliminate the risk of memory conflicts caused by fixed node combinations, while using a computational load splitting mechanism to balance the computing power differences between heterogeneous nodes and avoid task backlog. For example, in a prediction period with an L1 timeframe of 15 minutes, the computational load deviation for each node does not exceed 5% of the total task load.

[0025] Step S3: Analyze the rationality of task allocation in the computing center: At the end of the prediction period, the total time for each reserved object to execute computing tasks is counted and marked as the processing value of the reserved object. The variance of the processing values ​​of all reserved objects within the prediction period is calculated to obtain the distribution coefficient. At the end of the prediction period, the peak resource demand of all prediction periods is counted. The maximum and minimum values ​​of the peak resource demand constitute the resource demand range. The resource demand range is divided into several resource demand intervals. The reserved processing clusters of the prediction periods whose peak resource demand is within the resource demand intervals are marked as the matching clusters of the resource demand intervals. The matching cluster with the smallest distribution coefficient is marked as the configuration cluster of the resource demand interval. When the peak resource demand is obtained in the prediction period of the next prediction period, the configuration cluster of the resource demand interval corresponding to the peak resource demand is retrieved as the reserved processing cluster. The statistical object of the processing value is the total time for each computing node in the reserved processing cluster to execute computing tasks. This total time directly reflects the actual load intensity of the nodes during the predicted period. Variance calculation uses statistical methods to quantify the deviation of each processing value from the mean. The larger the variance value, the more uneven the task distribution among the nodes. The distribution coefficient, as a result of the variance calculation, is negatively correlated with the balance of task distribution. When the distribution coefficient approaches zero, it indicates that the task execution time distribution among the nodes in the cluster is uniform.

[0026] Specifically, at the end of the prediction period, the system iterates through the task execution records of all computing nodes within the reserved processing cluster, accumulating the total processing time of each node during this period as the processing value. For example, if a node executes three tasks sequentially during the prediction period, its processing value is the sum of the execution times of the three tasks. The system inputs the processing values ​​of each node into the variance calculation module, and calculates the variance value using a mathematical formula. This value is the distribution coefficient. After the distribution coefficient is generated, the system stores it along with historical data for use in optimizing cluster configuration selection in subsequent cycles. Through variance calculation, the system can objectively quantify the degree of load difference among nodes within the cluster, providing data support for selecting the configuration cluster with the smallest distribution coefficient in subsequent cycles, thereby ensuring that the task allocation scheme achieves optimal load balancing.

[0027] Step S4: Analyze the task execution status of the computing center: At the end of the prediction period, sum and average the processing values ​​of all reserved objects in the configuration cluster corresponding to the resource demand interval to obtain the execution coefficient of the resource demand interval. Compare the execution coefficient with the preset execution threshold: If the execution coefficient is less than the execution threshold, it is determined that the task execution status of the configuration cluster in the resource demand interval meets the requirements; if the execution coefficient is greater than or equal to the execution threshold, it is determined that the task execution status of the configuration cluster in the resource demand interval does not meet the requirements, and the corresponding resource demand interval is marked as the expanded prediction interval. Perform execution optimization analysis on the expanded prediction interval: Mark the number of times the allocated object is included in the reserved processing cluster in the current prediction period as the inclusion value. When the peak resource demand in the prediction period of the next prediction period is within the expanded prediction interval, obtain the allocation update value FPg through the formula FPg=[t1×FP], where t1 is the proportional coefficient, and 1.15≤t1≤1.25, and FP is the value of the allocation quantity. Add the allocated object with the largest inclusion value (excluding the reserved objects in the configuration cluster corresponding to the expanded prediction interval) to the configuration cluster until the number of reserved objects in the configuration cluster reaches the allocation update value FPg.

[0028] The processing value is defined as the total time for each reserved object to execute its computational task. The execution coefficient is obtained by calculating the arithmetic mean of the processing values ​​of all reserved objects. The execution threshold is preset to a fixed value to quantify the pass / fail standard for evaluating task execution efficiency. When the execution coefficient exceeds this threshold, the corresponding resource demand interval is marked as the expanded pre-reservation interval, triggering the optimization analysis process. During the optimization analysis, the inclusion value of the allocated object is recorded as the number of times it is selected into the reserved processing cluster within the current prediction period. In the next prediction period, when the peak resource demand is located within the expanded pre-reservation interval, the allocation update value is calculated by multiplying the original allocation quantity by a proportional coefficient, and the configuration cluster size corresponding to the expanded pre-reservation interval is dynamically expanded.

[0029] Specifically, at the end of the prediction period, the processing values ​​of all reserved objects are aggregated and averaged to generate an execution coefficient for the resource demand interval. This coefficient is compared with a preset threshold. If the execution coefficient exceeds the threshold, it indicates that the task execution efficiency of that interval has not met the standard, and the corresponding resource demand interval is marked as an expanded prediction interval. In the next prediction period, when the predicted peak resource demand is located in the expanded prediction interval, the allocation quantity is increased proportionally. The updated allocation value is calculated as the original allocation quantity multiplied by a proportional coefficient within the range of 1.15 to 1.25. Among the allocated objects not selected for the configuration cluster, the node with the highest inclusion value is preferentially added to the configuration cluster until the cluster size reaches the updated allocation value. By dynamically expanding the configuration cluster size, task processing capacity is improved, and task execution efficiency is effectively optimized.

[0030] The scaling factor t1 is limited to a range of 1.15 to 1.25, a range verified through experiments to balance resource expansion needs with cost control. The allocation update value FPg is calculated using a rounding-up method to ensure the number of reserved objects meets incremental requirements. The inclusion value is statistically based on the frequency with which allocated objects were selected as reserved objects within historical periods; high-frequency objects indicate superior processing capacity or stability. New reserved objects are selected preferentially from non-currently configured clusters, choosing the allocation object with the highest inclusion value to avoid local resource overload.

[0031] Specifically, when the execution coefficient exceeds a preset threshold, the system determines that the current cluster configuration cannot meet the task execution efficiency requirements and triggers the expansion interval optimization process. The allocation update value FPg is dynamically expanded based on the original allocated quantity FP using a scaling factor t1. For example, when FP is 100 and t1 is 1.2, FPg is calculated to be 120. Newly added reserved objects are selected from the allocated objects of non-currently configured clusters, prioritizing the nodes with the highest inclusion values. This mechanism dynamically optimizes the cluster size through historical data, improving task processing capabilities while avoiding resource idleness caused by excessive expansion. The inclusion value selection mechanism ensures that newly added nodes have high task processing experience, reducing the risk of load imbalance caused by differences in node performance.

[0032] The calculation of the allocation update value involves linearly expanding the original allocation quantity using a scaling factor t1, with the scaling factor limited to a range of 1.15 to 1.25. This ensures that the resource expansion can alleviate task backlog while avoiding over-allocation. The inclusion value serves as a quantitative indicator of the historical load status of the allocation object, reflecting the frequency with which the node participates in task processing within the historical period. When the peak resource demand corresponding to the expansion interval occurs, the allocation object with the largest inclusion value is preferentially selected to join the configuration cluster, leveraging its historical load experience to improve cluster stability.

[0033] Specifically, when the peak resource demand during the forecast period falls within the expansion interval, the allocation update value FPg is first calculated based on the ratio coefficient t1. For example, when FP is 100 and t1 is 1.2, FPg is updated to 120. Then, the allocation objects not selected into the current configuration cluster are traversed, and the node with the highest inclusion value is selected. For example, if a node was included in the reserved processing cluster 15 times in the previous forecast period, its inclusion value is higher than other nodes. The selected nodes are then added to the configuration cluster sequentially until the cluster size reaches FPg. During this process, the cumulative statistics of inclusion values ​​ensure that newly added nodes have experience in high-frequency task processing, while the range constraint of the ratio coefficient prevents the resource allocation from exceeding actual needs. For example, when t1 is set to 1.25, the upper limit of resource expansion is 25% of the original allocation, avoiding resource idleness due to over-allocation. Thus, the resource allocation within the expansion interval is dynamically adjusted, and new nodes are added preferentially based on historical load status, effectively alleviating task execution delays caused by insufficient resources or uneven node computing power.

[0034] This AI-based dynamic resource planning method for intelligent computing centers first performs periodic load forecasting analysis. A forecasting period is generated and divided into several forecasting time periods. At the beginning of each forecasting time period, historical task queue data, historical resource utilization data, and metadata are acquired. This data is input into a recurrent neural network for load analysis to obtain the peak resource demand and allocation quantity for the forecasting time period. Next, tasks are allocated to the computing nodes in the computing center. Reserved objects are randomly selected and these objects form a reserved processing cluster. Then, the rationality of the task allocation in the computing center is analyzed. At the end of the forecasting time period, the distribution coefficient of the reserved processing cluster is obtained. At the end of the forecasting period, the peak resource demand for all forecasting time periods is statistically analyzed, and the maximum and minimum values ​​of the peak resource demand constitute the resource demand range. The resource demand range is divided into several resource demand intervals, and the configuration clusters within each interval are marked. When the peak resource demand is obtained in the forecasting time period of the next forecasting period, the configuration clusters within the corresponding resource demand interval are retrieved as reserved processing clusters.

[0035] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0036] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0037] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for dynamic resource planning in intelligent computing centers based on artificial intelligence prediction, characterized in that, Includes the following steps: Step S1: Perform periodic predictive analysis on the load of the computing center: Generate a prediction period and divide the prediction period into several prediction periods. At the beginning of each prediction period, acquire historical data of the task queue, historical data of resource utilization, and metadata. Align and aggregate the historical data of the task queue, historical data of resource utilization, and metadata according to a uniform time frequency. Input the historical data of the task queue, historical data of resource utilization, and metadata into a recurrent neural network for load analysis to obtain the peak resource demand and allocation quantity for the prediction period. Step S2: Assign tasks to the computing nodes in the computing center: randomly select reserved objects and form a reserved processing cluster from the reserved objects; Step S3: Analyze the rationality of task allocation in the computing center: At the end of the prediction period, obtain the distribution coefficient of the reserved processing cluster; at the end of the prediction period, count the peak resource demand for all prediction periods; the maximum and minimum values ​​of the peak resource demand constitute the resource demand range; divide the resource demand range into several resource demand intervals; mark the configuration clusters in the resource demand intervals; when the peak resource demand is obtained in the prediction period of the next prediction period, retrieve the configuration clusters in the resource demand interval corresponding to the peak resource demand as the reserved processing clusters. Step S4: Analyze the task execution status of the computing center.

2. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 1, characterized in that, In step S1, the historical data of the task queue is high-frequency time-series data collected from the scheduler. The task queue data includes timestamps, the number of submitted tasks, the number of tasks in the queue, and the number of tasks running. Historical resource utilization data is collected from the monitoring system. This data includes total CPU utilization, memory usage, and GPU utilization. Metadata includes average task duration and requested CPU / GPU / memory.

3. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 2, characterized in that, In step S2, the specific process of task allocation for the computing nodes of the computing center includes: marking the computing nodes of the computing center as allocation objects, randomly selecting allocation objects as reserved objects, the number of reserved objects being equal to the number of allocation objects, forming a reserved processing cluster from the reserved objects, and evenly distributing the computing tasks received during the prediction period to the reserved objects in the reserved processing cluster for processing according to the amount of computing.

4. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 3, characterized in that, In step S3, the process of obtaining the uniformity coefficient includes: counting the total time for each reserved object to perform the calculation task and marking it as the processing value of the reserved object; and calculating the variance of the processing values ​​of all reserved objects within the prediction period to obtain the uniformity coefficient.

5. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 4, characterized in that, In step S3, the specific marking process for the configuration clusters of the resource demand range includes: marking the reserved processing clusters of the prediction period whose resource demand peak is within the resource demand range as the matching clusters of the resource demand range, and marking the matching cluster with the smallest distribution coefficient as the configuration clusters of the resource demand range.

6. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 5, characterized in that, In step S4, the specific process of analyzing the task execution status of the computing center includes: at the end of the prediction period, summing and averaging the processing values ​​of all reserved objects in the configuration cluster corresponding to the resource demand interval to obtain the execution coefficient of the resource demand interval, and judging whether the task execution status of the configuration cluster in the resource demand interval meets the requirements by using the execution coefficient. If the requirements are not met, the execution optimization analysis is performed on the extended pre-determined interval.

7. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 6, characterized in that, The specific process for determining whether the task execution status of the configuration cluster within the resource demand range meets the requirements includes: comparing the execution coefficient with a preset execution threshold; if the execution coefficient is less than the execution threshold, the task execution status of the configuration cluster within the resource demand range is determined to meet the requirements; if the execution coefficient is greater than or equal to the execution threshold, the task execution status of the configuration cluster within the resource demand range is determined to not meet the requirements, and the corresponding resource demand range is marked as an extended range.

8. The method for dynamic resource planning of intelligent computing centers based on artificial intelligence prediction according to claim 7, characterized in that, The specific process of performing optimization analysis on the extended forecast interval includes: marking the number of times the allocated object is included in the reserved processing cluster within the current forecast period as the inclusion value; when the peak resource demand during the forecast period in the next forecast period is within the extended forecast interval, the allocation update value FPg is obtained through the formula FPg=[t1×FP], where t1 is the proportional coefficient and 1.15≤t1≤1.25, and FP is the value of the allocation quantity; the allocated object with the largest inclusion value other than the reserved object in the configuration cluster corresponding to the extended forecast interval is added to the configuration cluster until the number of reserved objects in the configuration cluster reaches the allocation update value FPg.

Images