Intelligent information system resource control method and system based on large model

Abstract

The invention belongs to the technical field of information system resource control, and particularly discloses an intelligent information system resource control method and system based on a large model, and the method comprises the steps: extracting the text data of a request intention, a task type and a task priority label based on a user request log; extracting external event text data based on the external environment data; splicing the index vector and the text vector to obtain an input matrix; predicting the input matrix by using the large model to obtain a task-level resource demand map and a resource control instruction sequence; and taking the task-level resource demand, the resource control instruction sequence, the current resource state of the system, the dependency relationship between the tasks and the historical performance indexes of the system obtained by the large model as input, and obtaining a final task queue priority ranking result and a resource allocation result through a reinforcement learning algorithm. The large model can continuously learn and improve the prediction accuracy and the effectiveness of the control strategy, and self-adaptive optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of information system resource control technology, and in particular to an intelligent information system resource control method and system based on a large model. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of big data and cloud computing, the complexity of intelligent information systems and the amount of data involved are constantly increasing, posing a huge challenge to the resource management of information systems.

[0004] Resource management of information systems mainly includes performance monitoring and capacity management of CPU, memory, disk I / O, network bandwidth, etc., as well as managing components on which the information system depends, such as the search engine Elasticsearch, the cache Redis, and the database MySQL. Currently, most resource management for information systems adopts control methods based on fixed rules or thresholds. For example, the system presets fixed thresholds for CPU utilization, memory usage, and request queue length. When monitored indicators exceed or fall below a certain threshold, predefined scaling-up or scaling-down actions are triggered (such as adding or removing Pod / container instances, adjusting virtual machine specifications, allocating or releasing bandwidth, etc.). However, this approach cannot effectively cope with complex, variable, and non-linear load patterns (such as sudden traffic spikes, periodic peaks, and sudden changes in user behavior). Furthermore, resource adjustments occur after load changes, which may lead to service degradation (scaling-up lag) or resource idleness (scaling-down lag). Moreover, preset thresholds often require experienced operations personnel to repeatedly adjust them, making it difficult to achieve optimal performance.

[0005] Existing technologies also disclose control methods based on simple statistics or time series forecasting. These methods use historical data and simple time series models such as moving averages, exponential smoothing, or ARIMA to predict short-term load trends and adjust resources accordingly. Some systems also attempt to predict load or resource requirements using supervised learning models (such as linear regression, SVM, decision trees, and shallow neural networks). However, these methods are insufficient when dealing with high-dimensional, heterogeneous, and context-dependent monitoring data (such as combinations of multiple indicators, log semantics, and external events), making it difficult to accurately predict future resource requirements, especially lacking foresight regarding "black swan" events or new patterns.

[0006] Furthermore, the aforementioned methods primarily rely on system-level metrics (such as CPU, memory, and network I / O), severely lacking an understanding and utilization of higher-level contextual information such as business semantics, user intent, and task type. This makes it difficult to perform refined and differentiated resource allocation and priority control based on the specific needs of the task. For example, they cannot distinguish the differences in resource requirements between high-priority and low-priority tasks, or dynamically adjust allocation strategies based on task content (such as complex queries vs. simple queries). Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a resource control method and system for intelligent information systems based on a large model. The method constructs an intelligent adaptive resource control hub for information systems based on the large model, using structured indicators and unstructured semantics as inputs. Through multi-dimensional semantic reasoning, it can predict the refined demand distribution of multiple resource dimensions at the task level.

[0008] In some implementations, the following technical solutions are adopted: A resource control method for intelligent information systems based on a large model, comprising: Real-time acquisition of system monitoring metrics, user request logs, and external environment data; Extract text data of request intent, task type, and task priority tags from user request logs; extract text data of external events from external environment data; The acquired numerical indicator data is normalized to generate indicator vectors; the acquired text data is vectorized to generate semantic vectors; the indicator vectors and text vectors are concatenated to obtain the input matrix. The large model is used to predict the input matrix to obtain a task-level resource requirement map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependencies, and the expected resource consumption of the task. Using the task-level resource requirements, resource control instruction sequences, current system resource status, dependencies between tasks, and historical system performance metrics obtained from the large model as input, the final priority ranking result of the task queue and the resource allocation result are obtained through reinforcement learning algorithm.

[0009] As a further measure, the system monitoring metrics include: CPU utilization, memory usage, network I / O, request queue length, and the operational status of dependent components.

[0010] As a further solution, a set input prompt word template is used to guide the large model inference. The input prompt word template is as follows: when the current system state is {index vector}, the business scenario is {semantic description}, and the external event is {event summary}, it is necessary to predict the resource requirements of various types within the future {time} and generate a sequence of resource control instructions, with the optimization objective being {maximizing throughput}.

[0011] As a further embodiment, the reward function of the reinforcement learning algorithm includes: Rewards for successful tasks, penalties for delays, rewards for resource utilization, priority rewards for critical tasks, penalties for resource waste, weighted rewards based on semantic importance, and rewards based on task dependency chains. The final reward function is obtained by weighted summation of all reward and penalty items.

[0012] As a further option, the adaptive adjustment strategy for the weights of each reward and penalty item can be determined as follows: ; in, The basic weights for rewards and penalties are set according to the actual situation; This indicates the real-time parameter value corresponding to the reward or penalty item. For the corresponding parameter threshold; To control sensitivity, To control the decay rate, This is the time interval since the last adjustment.

[0013] As a further solution, it also includes: calculating the resource deviation index, the service level agreement dynamic compliance rate index, and the resource elasticity response index respectively, and triggering adjustments to the output of the large model and the weights of the reward function in the reinforcement learning algorithm based on the calculation results of the indexes.

[0014] As a further option, the resource deviation indicators include: resource utilization deviation, cost deviation rate, and network latency deviation; The dynamic compliance rate indicators for the service level agreement include: SLA compliance rate and delay compliance rate. The resource resilience response index includes: resilience response index and resource recovery timeliness.

[0015] As a further solution, adjustments to the output of the large model can be triggered based on the calculation results of the indicators, specifically: If the result of the indicator calculation contradicts the resource allocation strategy output by the large model, the result of the large model output will be automatically adjusted to match the result of the indicator calculation. And / or, the calculation results of the indicators are fed back to the large model as input, so that it automatically considers the current system state when generating the next resource allocation strategy.

[0016] In other embodiments, the following technical solutions are adopted: A resource control system for intelligent information systems based on a large model, comprising: The data acquisition module is configured to acquire system monitoring metrics, user request logs, and external environment data in real time. The data extraction module is configured to extract text data based on user request logs, including request intent, task type, and task priority tags; and to extract text data of external events based on external environment data. The feature generation module is configured to normalize the acquired numerical index data to generate index vectors; vectorize the acquired text data to generate semantic vectors; and concatenate the index vectors and text vectors to obtain an input matrix. The resource prediction module is configured to use a large model to predict the input matrix to obtain a task-level resource demand map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependencies, and the expected resource consumption of the task. The scheduling optimization module is configured to take the task-level resource requirements obtained from the large model, the resource control instruction sequence, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as inputs, and obtain the final priority ranking result of the task queue and the resource allocation result through reinforcement learning algorithm.

[0017] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, wherein the processor implements instructions; and the memory stores multiple instructions adapted to be loaded and executed by the processor in accordance with the above-described resource control method for intelligent information systems based on large models.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention can predict traffic surges in advance by parsing social media text, proactively warming up container instances and avoiding service downtime. By fusing structured system monitoring metrics (such as CPU, memory, and network) with unstructured semantic information (such as user request content, task type descriptions, log text summaries, and external event text), and inputting them into a large language model (LLM) or a large multimodal model (LMM), the powerful natural language understanding and context modeling capabilities of the large model are utilized to achieve a deep understanding of the information system's operating status, task attributes, and external environment. Based on the fused heterogeneous context information, a refined demand map of different types of tasks / requests for various resource types (including but not limited to CPU, GPU, memory, storage IO, and network bandwidth) is predicted within a specific future time window.

[0019] Meanwhile, based on the large model's understanding of the context and predictions of future resource demands, combined with preset optimization objectives, the large model directly generates specific, fine-grained sequences of resource control instructions. These instructions perform differentiated operations (such as reservation, restriction, migration, scaling, warm-up, and recycling) for specific tasks, services, or resource units, rather than simply expanding or shrinking the overall resource pool.

[0020] (2) This invention takes task feature vectors, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as inputs, and obtains the final priority ranking result of the task queue and the resource allocation result through reinforcement learning algorithm; under the condition of limited resources, it intelligently schedules different tasks to maximize system throughput. By setting different reward and punishment items and dynamically adjusting the weight of each reward and punishment item, the model can make timely adaptive adjustments when facing sudden traffic scenarios, and can effectively cope with complex, variable, and nonlinear load patterns.

[0021] (3) This invention continuously collects actual effect data after the execution of resource control instructions, and reflects the predictive performance of the model through the constructed large model evaluation index; it uses the evaluation index to adjust the weight of the large model parameters and reward function, and enables the large model to continuously learn and improve its predictive accuracy and the effectiveness of the control strategy through online fine-tuning, reinforcement learning strategy optimization or prompting engineering adjustment, so as to achieve adaptive optimization. Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of a resource control method for an intelligent information system based on a large model, as described in an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Example 1 In one or more embodiments, a resource control method for intelligent information systems based on a large model is disclosed, combining... Figure 1 Specifically, it includes the following process: S101: Real-time acquisition of system monitoring metrics, user request logs, and external environment data.

[0026] In this embodiment, system monitoring metrics data are collected. These metrics include GPU and CPU utilization, memory usage, network I / O, request queue length, and the running status of dependent third-party components, such as whether the database is running normally.

[0027] Data is collected at a set sampling frequency, such as once per second, and historical data from the most recent 24 hours is retained.

[0028] User request logs can be obtained by accessing the access log files generated by web servers (such as Nginx and Apache). User request logs are a key component in system monitoring and operation and maintenance analysis. They record detailed information about every request a user makes to the system. These mainly include: basic request information (such as timestamp, client IP address, request URL / path, request protocol, response status code, etc.), request details and context (such as user's source, identity, browser, operating system, device model, request data type, data volume, request parameters, etc.), and server response information (such as the amount of data returned by the server to the client, the total time the server spent processing the request, and the hardware resources required to process the request).

[0029] External environment data refers to textual information crawled from social media or news APIs (application programming interfaces), such as real-time or historical news articles, summaries, sources, and other information.

[0030] S102: Extract text data of request intent, task type, and task priority tags based on user request logs; extract text data of external events based on external environment data.

[0031] In this embodiment, the pre-trained language model BERT is used to analyze and process user request logs, extracting text data of request intent (such as "real-time transaction" or "data backup"), task type (such as "video transcoding" or "database query"), and task priority tags (matched by keywords, such as "VIP" marked as the highest level).

[0032] In this embodiment, an event extraction model is used to parse external event text and identify key information that may affect system resources, such as promotional activities and specific time points, and trending topics on Weibo.

[0033] Event extraction refers to identifying dynamically occurring events (such as promotions, disasters, transactions, etc.) from text and extracting key elements related to the events (such as time, location, participants, trigger words, etc.).

[0034] Event extraction models can be implemented by directly using existing open-source large models in conjunction with prompt word projects, or by pre-training existing open-source large models and then fine-tuning them to achieve end-to-end extraction.

[0035] The input to the event extraction model is the crawled text information, such as "Double 11 promotion starts in 24 hours". The event extraction model will identify "promotion" and the time element "24 hours later", and then organize the information into key-value pairs in JSON format for structured output. The event extraction model can transform natural language text into machine-processable structured data, providing a foundation for subsequent predictive analysis and decision support.

[0036] S103: Normalize the acquired numerical index data to generate index vectors; vectorize the acquired text data to generate semantic vectors; concatenate the index vectors and text vectors to obtain the input matrix.

[0037] In this embodiment, numerical index data undergoes normalization processing, for example: CPU utilization = (current value - minimum value) / (maximum value - minimum value).

[0038] All normalized indicator data form a numerical indicator vector.

[0039] For text-based data, the Sentence-BERT model can be used to generate fixed-dimensional semantic vectors (e.g., 768-dimensional) from text data containing request intent, task type, priority labels, and semi-structured text data obtained from the event extraction model. The Sentence-BERT model is an existing model structure that can encode each part of the text separately and then concatenate them into a complete semantic vector; the specific encoding method will not be detailed here.

[0040] The obtained numerical index vector and semantic vector are concatenated into an input matrix of a unified format [numerical index vector || semantic vector], which is used to adapt to the input requirements of large models.

[0041] S104: Use a large model to predict the input matrix to obtain a task-level resource requirement map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependency relationship (i.e. whether the execution of a task depends on the result of other tasks) and the expected resource consumption of the task.

[0042] In this embodiment, the large model can be an existing large language model (LLM) or large multimodal model (LMM); the large model can be fine-tuned using historical resource allocation records and task-level business scenario annotations (including request intent, task type, task priority, and external events).

[0043] The input matrix, which includes the current system monitoring metrics, user request data, and information on the impact of external events, is used as the input to the large model. Based on these inputs, specific input prompt word templates are used to guide the large model's reasoning. For example: "The current system state is {metric vector}, the business scenario is {semantic description}, the external event is {event summary}, we need to predict the resource requirements of various types in the next 10 minutes and generate control strategies, and the optimization objective is {maximize throughput (the number of requests that can be processed)}." The large model performs multi-dimensional predictions based on the received information to obtain a task-level resource demand map. This allows for the prediction of refined demand distributions across multiple resource dimensions, such as GPU and CPU utilization, memory usage, network I / O, and request queue length, at the task level. Each resource dimension has a time series representation, indicating the changes in that resource during task execution. The output structure can be JSON, DataFrame, or tensor. For example: {"task_id":"inference-bert-32batch","resource_demand_graph" / / Task-level resource demand graph:} {"gpu_utilization":[65,72,78,70,68], / / GPU utilization (%) over 5 time slices; "cpu_utilization":[20,25,30,28,22], / / CPU utilization (%); "memory_usage":[8192,9216,10240,9728,8704], / / Memory usage (MB); "network_io":[12.4,15.2,14.8,13.6,11.9], / / Network I / O (MB / s); "request_queue_length":[2,4,6,3,1] / / Request queue length}; "time_resolution": "200ms", / / Granularity of each time slice; "predicted_duration": "1.0s" / / Total task time}.

[0044] Based on prediction results and contextual understanding, specific resource control strategy instructions are dynamically generated; for example: The "Preheat" operation targets the GPU resource pool and specifies 3 GPU units. That is, when a user request is detected to contain the keyword "real-time video analysis", 3 GPU resources are allocated to the task in advance. Before the actual use, 3 GPU resources are activated or allocated in advance to put them into standby state, so as to better meet the requirements of real-time video analysis.

[0045] The "Limit" operation sets the maximum CPU usage of the batch task named batch-job-001 to 0.5 cores; that is, it prevents the batch task named batch-job-001 from consuming more than 50% of the single-core CPU resources, avoiding exhaustion of system resources or affecting other tasks; it limits the CPU usage of low-priority tasks in a shared server, thereby ensuring system stability and fair allocation of resources.

[0046] The "Prioritize" operation targets VIP users, increasing their network bandwidth by 200Mbps. This means allocating higher network transmission priority and additional bandwidth to VIP users to ensure their data transmission speed and stability. In the event of network congestion, the needs of VIP users for video conferencing, real-time data transmission, and other services can be prioritized.

[0047] After receiving these task-level resource control policy instructions, the large model, under the premise of limited resources, needs to further combine multiple tasks that need to run in the system (such as tasks obtained based on user requests, data inference tasks, cache update tasks, etc.) and use reinforcement learning methods to achieve dynamic task priority evaluation. The specific process is explained in detail in step S105.

[0048] S105: Taking the task-level resource requirements, resource control instruction sequence, current system resource status, dependencies between tasks, and historical system performance indicators obtained from the large model as input, the final task queue priority ranking result and resource allocation result are obtained through reinforcement learning algorithm.

[0049] Specifically, the current resource status of the system describes the available resources at a given moment, typically including: computing resources (such as GPU or CPU utilization, number of available threads, etc.), memory resources (total memory, used memory, available memory, memory bandwidth usage, etc.), storage resources (such as disk I / O throughput, disk space, etc.), network resources (such as network bandwidth utilization, network latency), task queue status (such as the number of currently running tasks, the number of tasks waiting to be executed, the average waiting time of tasks in the queue, the task scheduler load, etc.), and resource allocation status (such as the resource usage and resource fragmentation of each node / container, etc.).

[0050] Historical performance metrics of the system are used by reinforcement learning models to learn scheduling strategies and evaluate scheduling effectiveness; these mainly include: task execution performance, resource utilization efficiency, system throughput, quality of service (QoS) metrics, and historical feedback on scheduling strategies.

[0051] This embodiment uses the predicted task-level resource requirements, resource control instruction sequences, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as state variables for reinforcement learning. The execution actions are to prioritize each task and allocate resources to each task. At the same time, a reward function is set, and the execution effect of the strategy is fed back by evaluating the execution results. Then, the reward function is updated, and through continuous optimization of decisions, the final priority ranking result of the task queue and the resource allocation result are obtained.

[0052] In this embodiment, the reward function specifically includes: (1) Rewards for successful task completion (positive items): A fixed positive reward is given when the task is completed within the specified time: ; The reward can also be combined with task priority; for example, the reward for a successful high-priority task can be set to 1.5.

[0053] (2) Delay penalty (negative item): For tasks that time out, the penalty increases with the delay time:

[0054] If the task needs to be completed within 1 hour, 0.05 will be deducted for every 1 minute of delay, and 0.5 will be deducted for every 10 minutes of delay.

[0055] (3) Resource utilization rate bonus (positive item): The system is encouraged to utilize resources efficiently while completing tasks: ; If the average utilization of GPU and CPU reaches 90%, the reward value is 0.2 × 0.9 = 0.18.

[0056] (4) Priority reward for critical tasks (positive item): Additional rewards are given for prioritizing the scheduling of high-priority tasks (such as real-time transactions and video conferencing): ; For example, this reward can be triggered when a VIP user prioritizes their bandwidth.

[0057] (5) Penalties for resource waste (negative items): Penalize over-allocated but unused resources: ; If 500MB of memory is allocated but only 300MB is used, and the penalty unit for resource waste is 100MB, then the penalty value is -0.1×2= -0.2.

[0058] (6) Semantic importance-weighted reward: By analyzing the semantics of tasks using a large model, different weights can be assigned to different tasks. For example, the weight of user complaint handling can be set to 1.5 times that of system log analysis tasks.

[0059] (7) Task dependency chain rewards: The scheduling strategy rewards tasks that depend on others and execute them ahead of schedule. For example, if task B depends on task A, and A is executed ahead of schedule, the reward is +0.2, which promotes the efficiency of system collaboration.

[0060] The semantic importance-weighted reward in this embodiment can improve the system's ability to understand and respond to business priorities; the task dependency chain reward can improve the system's ability to optimize task coordination and overall link efficiency.

[0061] We obtain the final reward function by weighted summation of all reward and penalty factors: ; in, This is represented as the i-th type of reward or punishment item. This represents the weight corresponding to each type of reward or punishment item; n represents the number of reward or punishment items.

[0062] The initial weights of the above-mentioned reward and punishment items are determined through ablation experiments, and can also be adjusted according to actual business needs.

[0063] In this embodiment, the weight adaptive adjustment strategy enables the system to effectively cope with complex, variable, and nonlinear load patterns (such as sudden traffic surges, periodic spikes, and sudden changes in user behavior), thereby improving the system's decision-making accuracy.

[0064] As a specific implementation method, the adaptive adjustment strategy for the weights of each reward and penalty item can be determined as follows: ; in, The basic weights for rewards and penalties are set according to the actual situation; This indicates the real-time parameter value (which can be the actual quantity or a ratio) corresponding to the reward or penalty item. For the corresponding parameter threshold; To control sensitivity, To control the decay rate, This is the time interval since the last adjustment.

[0065] This method uses a continuous function to adaptively calculate the weights of each reward and punishment item. The weights can change smoothly with the indicators, and the adjustment range is automatically decayed through an exponential decay factor to prevent over-adjustment.

[0066] As a specific example, the weight for the delay penalty is calculated as follows:

[0067] in, As the base weight for delayed penalties, Indicates the real-time latency rate. Indicates the baseline delay threshold. To control sensitivity, To control the decay rate, This is the time interval since the last adjustment.

[0068] In other real-time methods, the weights of each reward and penalty item can be dynamically adjusted using system operation data. Specifically, the system throughput, latency indicators, and historical weights of each reward and penalty item are used as inputs to the PPO (Proximal Policy Optimization) algorithm, which automatically outputs a weight allocation scheme for each reward and penalty item.

[0069] At the same time, it allows intervention in weight adjustments through preset rules; for example: If the SLA compliance rate for a certain service is less than 95%, the weight of the mandatory task success reward will be increased to 1.5.

[0070] During periods of tight resource budgets, the adjustment range of resource utilization reward weights is automatically frozen.

[0071] Of course, the weight of each reward and punishment item can be adjusted through subsequent evaluation indicators, which will be explained later.

[0072] By employing a dynamic weight adjustment strategy, the SLA (Service Level Agreement) compliance rate can still be maintained above 99% when facing sudden traffic surges, such as a 300% increase in QPS (queries per second).

[0073] In this embodiment, after obtaining the priority and resource allocation results of the large model output, task queue scheduling (such as priority queue or preemptive scheduling), GPU or CPU resource scheduling, memory or cache management, and network bandwidth allocation are performed based on the results.

[0074] Then, based on the task execution, evaluation indicators are calculated. Specific evaluation indicators mainly include: (1) Resource deviation indicators, including resource utilization deviation, cost deviation rate and network latency deviation; Resource utilization deviation = Actual utilization rate - Predicted utilization rate; Cost deviation rate = ; ; in, This refers to the maximum acceptable network latency value promised by the service provider; for example, if a cloud service provider promises "network latency not exceeding 50ms", then the SLA threshold is 50ms.

[0075] The aforementioned resource deviation index quantifies the model's processing effectiveness of contextual information such as business semantics and user intent by comparing predicted values ​​with actual execution results.

[0076] (2) Service Level Agreement (SLA) dynamic compliance rate indicators, including: SLA compliance rate and delay compliance rate indicators; ; Delayed achievement rate = ; The Service Level Agreement (SLA) dynamic compliance rate metric can track the achievement of key service indicators in real time, ensuring the service quality of high-priority tasks. When the compliance rate of a certain type of task is consistently below 95%, the large model is triggered to re-evaluate the semantic feature weights (i.e., the weights of the semantic importance-weighted reward) of that type of task. If the compliance rate is <90% for three consecutive periods, it will automatically roll back to the best historical strategy version.

[0077] (3) Resource elasticity response index indicators include: elasticity response index and resource recovery timeliness.

[0078] Elastic response index = ; Timeliness of resource recycling = ; The resource elasticity response index can quantify the latency of resource reclamation and expansion, ensuring rapid response in scenarios with sudden traffic surges. When this index is continuously below 10% of the benchmark value, it indicates that the current resource allocation strategy may be lagging, and the system will automatically trigger dynamic optimization of the resource allocation strategy to better match the actual load demand. The resource reclamation timeliness index can effectively identify the lag in resource release, avoiding cost waste and decreased resource utilization caused by excessive resource occupation.

[0079] In this embodiment, based on the calculation results of the above indicators, adjustments to the output of the large model and the weights of the reward function in the reinforcement learning algorithm can be triggered.

[0080] As a specific example: (1) If the result of the indicator calculation is inconsistent with the resource allocation strategy output by the large model (or is not compatible), the result of the large model output will be automatically adjusted to match the result of the indicator calculation. For example, if the current network latency is high, but the large model suggests using remote resources, the use of remote resources will be automatically adjusted to local resource scheduling.

[0081] (2) The calculation results of the indicators are fed back to the large model as input, so that it can automatically consider the current system status when generating the next resource allocation strategy; for example, if the CPU utilization is detected to be too high (>80%), the large model is prompted that the current resources are tight and resource release or task scheduling optimization is given priority. If the task completion rate decreases, the output results of the large model are guided to prioritize the execution of core tasks.

[0082] Furthermore, the calculation results of the indicators can also trigger the adjustment of the weights of the reward function in the reinforcement learning algorithm, allowing the reinforcement learning process to automatically adapt to the current system bottlenecks; for example, if the current memory usage is too high, the weight of "resource waste penalty" will be increased; if the task completion rate decreases, the weight of "task success reward" will be increased.

[0083] This embodiment takes task feature vectors, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as inputs. It uses a reinforcement learning algorithm to obtain the final priority ranking of the task queue and the resource allocation results. Under limited resource conditions, it intelligently schedules different tasks to maximize system throughput. By setting different reward and penalty items and dynamically adjusting the weights of each item, the model can make timely adaptive adjustments when facing sudden traffic surges, effectively coping with complex, variable, and nonlinear load patterns.

[0084] Example 2 In one or more embodiments, a resource control system for an intelligent information system based on a large model is disclosed, comprising: The data acquisition module is configured to acquire system monitoring metrics, user request logs, and external environment data in real time. The data extraction module is configured to extract text data based on user request logs, including request intent, task type, and task priority tags; and to extract text data of external events based on external environment data. The feature generation module is configured to normalize the acquired numerical index data to generate index vectors; vectorize the acquired text data to generate semantic vectors; and concatenate the index vectors and text vectors to obtain an input matrix. The resource prediction module is configured to use a large model to predict the input matrix to obtain a task-level resource demand map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependencies, and the expected resource consumption of the task. The scheduling optimization module is configured to take the task-level resource requirements obtained from the large model, the resource control instruction sequence, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as inputs, and obtain the final priority ranking result of the task queue and the resource allocation result through reinforcement learning algorithm.

[0085] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0086] Example 3 In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the processor to implement the resource control method for intelligent information systems based on large models as described in Embodiment 1.

[0087] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0088] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0089] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0090] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A resource control method for intelligent information systems based on a large model, characterized in that, include: Real-time acquisition of system monitoring metrics, user request logs, and external environment data; Extract text data of request intent, task type, and task priority tags from user request logs; extract text data of external events from external environment data; The acquired numerical indicator data is normalized to generate indicator vectors; the acquired text data is vectorized to generate semantic vectors; the indicator vectors and text vectors are concatenated to obtain the input matrix. The large model is used to predict the input matrix to obtain a task-level resource requirement map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependencies, and the expected resource consumption of the task. Using the task-level resource requirements, resource control instruction sequences, current system resource status, dependencies between tasks, and historical system performance metrics obtained from the large model as input, the final task queue priority ranking and resource allocation results are obtained through reinforcement learning algorithms.

2. The resource control method for an intelligent information system based on a large model as described in claim 1, characterized in that, The system monitoring metrics include: CPU utilization, memory usage, network I / O, request queue length, and the operational status of dependent components.

3. The resource control method for an intelligent information system based on a large model as described in claim 1, characterized in that, The large model inference is guided by a set input prompt word template. Specifically, when the current system state is {index vector}, the business scenario is {semantic description}, and the external event is {event summary}, it is necessary to predict the resource requirements of various types within the future {time} and generate a sequence of resource control instructions. The optimization objective is {maximizing throughput}.

4. The resource control method for an intelligent information system based on a large model as described in claim 1, characterized in that, The reward function of the reinforcement learning algorithm includes: Rewards for successful tasks, penalties for delays, rewards for resource utilization, priority rewards for critical tasks, penalties for resource waste, weighted rewards based on semantic importance, and rewards based on task dependency chains. The final reward function is obtained by weighted summation of all reward and penalty items.

5. The resource control method for an intelligent information system based on a large model as described in claim 4, characterized in that, The adaptive adjustment strategy for the weights of each reward and penalty item can be determined as follows: ； in, The basic weights for rewards and penalties are set according to the actual situation; This indicates the real-time parameter value corresponding to the reward or penalty item. For the corresponding parameter threshold; To control sensitivity, To control the decay rate, This is the time interval since the last adjustment.

6. The resource control method for an intelligent information system based on a large model as described in claim 1, characterized in that, Also includes: The resource deviation index, the dynamic compliance rate of the service level agreement, and the resource elasticity response index are calculated separately. Based on the calculation results of the indicators, adjustments are triggered to the output of the large model and the weights of the reward function in the reinforcement learning algorithm.

7. The resource control method for an intelligent information system based on a large model as described in claim 6, characterized in that, The resource deviation indicators include: resource utilization deviation, cost deviation rate, and network latency deviation; The dynamic compliance rate indicators for service level agreements include: SLA compliance rate and delay compliance rate. The resource resilience response index includes: resilience response index and resource recovery timeliness.

8. The resource control method for an intelligent information system based on a large model as described in claim 6, characterized in that, The calculation results of the indicators trigger adjustments to the output of the large model, specifically: If the result of the indicator calculation contradicts the resource allocation strategy output by the large model, the result of the large model output will be automatically adjusted to match the result of the indicator calculation. And / or, the calculation results of the indicators are fed back to the large model as input, so that it automatically considers the current system state when generating the next resource allocation strategy.

9. A resource control system for an intelligent information system based on a large model, characterized in that, include: The data acquisition module is configured to acquire system monitoring metrics, user request logs, and external environment data in real time. The data extraction module is configured to extract text data based on user request logs, including request intent, task type, and task priority tags; and to extract text data of external events based on external environment data. The feature generation module is configured to normalize the acquired numerical index data to generate index vectors; vectorize the acquired text data to generate semantic vectors; and concatenate the index vectors and text vectors to obtain an input matrix. The resource prediction module is configured to use a large model to predict the input matrix to obtain a task-level resource demand map and a resource control instruction sequence; the resource control instruction sequence includes: the task type of each task in the task queue, the inter-task dependencies, and the expected resource consumption of the task. The scheduling optimization module is configured to take the task-level resource requirements obtained from the large model, the resource control instruction sequence, the current resource status of the system, the dependencies between tasks, and the historical performance indicators of the system as inputs, and obtain the final priority ranking result of the task queue and the resource allocation result through reinforcement learning algorithm.

10. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-8, for the resource control method of a large-model-based intelligent information system.

Images