Multi-layer virtual machine scheduling system based on large language model

Abstract

The invention discloses a multi-layer virtual machine scheduling system based on a large language model, and the system comprises the following modules: an initial heuristic design module which is used for designing an initial heuristic function of a virtual machine scheduling problem; the scheduling scene feature analysis module is used for extracting and analyzing features of the cloud scheduling scene; the scene self-adaptive optimization module is used for performing independent optimization according to different scenes; the multi-scene strategy management module is used for managing cross-scene strategy switching; and the performance evaluation module evaluates the optimization effect in real time. By combining the intelligent reasoning capability of the large language model, the system can provide an efficient, flexible and self-adaptive scheduling optimization scheme, and the utilization efficiency of cloud resources is remarkably improved. Through the advantages of the large language model in the aspect of code generation, a heuristic function adaptive to virtual machine scheduling can be intelligently and efficiently generated, so that resource allocation is optimized, the efficiency and performance of a high-dynamic cloud scheduling problem are remarkably improved, and an innovative solution is provided for cloud resource scheduling.

Description

Technical Field

[0001] The present invention designs a multi-layer virtual machine angle system, especially a multi-layer virtual machine scheduling system based on large language models, belonging to the technical field of virtual machine scheduling optimization systems. Background Art

[0002] With the rapid development of information technology, cloud computing has become an important cornerstone of today's digital transformation. Through virtualization technology, cloud computing realizes the efficient utilization and flexible management of computing resources. These virtualization platforms abstract physical resources into multiple independent virtual machines, providing strong support for various applications. In this process, virtual machine scheduling, cloud resource management, and the dynamic nature of virtual machines have become hotspots in research and practice, aiming to improve the performance, reliability, and scalability of cloud computing platforms.

[0003] Cloud computing is a computing model based on the Internet that provides scalable computing resources and services through the network, and users do not need to concern themselves with the specific implementation of the underlying hardware and infrastructure. Cloud computing is mainly divided into three modes: Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS) to meet different levels of needs. Its core advantages lie in the elastic allocation, efficient utilization, and pay-per-use of resources, which greatly promotes the popularization of information technology and application innovation.

[0004] Cloud resource management is the process of effectively allocating and scheduling computing, storage, network, and other resources in a cloud computing environment to meet user needs and optimize resource utilization. It covers multiple aspects such as resource monitoring, allocation, scheduling, and optimization. Resource monitoring provides data support for decision-making by collecting (for example, using tools such as Prometheus, Nagios, etc.) resource usage in real time; resource allocation reasonably allocates resources to each application and service based on user needs and resource status; resource scheduling involves task arrangement and priority determination to improve the overall system efficiency and response speed; resource optimization further improves resource utilization, reduces energy consumption, and operating costs through various algorithms and strategies (such as heuristic algorithms, reinforcement learning algorithms).

[0005] Virtual machine scheduling is a core part of cloud resource management. It refers to how to reasonably allocate physical resources (such as CPU, memory, storage, and network) to each virtual machine in a multi-virtual machine environment to achieve performance optimization and maximize resource utilization. Its goals usually include improving system throughput, reducing response time, balancing the load, and saving energy consumption. Common virtual machine scheduling algorithms include load-balancing-based scheduling, priority-based scheduling, and energy-efficiency-based scheduling, etc. Load-balancing scheduling aims to evenly distribute the load to prevent some physical nodes from being overloaded while others are idle; priority scheduling dynamically adjusts resource allocation according to the importance of the virtual machine and the urgency of the task; energy-efficiency scheduling focuses on minimizing energy consumption while meeting performance requirements, and achieves the energy-saving goal by dynamically adjusting the on-off state of resources.

[0006] In recent years, with the development of artificial intelligence and machine learning technologies, intelligent scheduling algorithms (such as algorithms based on deep learning and reinforcement learning) have gradually been applied to virtual machine scheduling. Through learning and predicting historical data, these algorithms can achieve more accurate and efficient resource allocation. For example, the deep learning-based scheduling algorithm can predict the change in resource requirements of virtual machines and adjust the resource allocation strategy in advance, improving the system's adaptability and response speed. In addition, the application of big data analysis technology enables scheduling algorithms to process and analyze massive resource usage data and optimize scheduling decisions.

[0007] The dynamic problem of virtual machines mainly refers to the dynamic changes in the resource requirements of virtual machines and the resulting scheduling and management challenges in the cloud computing environment. Due to the diversity and unpredictability of user requirements, the resource usage of virtual machines may fluctuate over time and with the application load, which poses higher requirements for cloud resource management and virtual machine scheduling. First of all, the dynamic nature of virtual machines increases the complexity of resource scheduling. The scheduling algorithm needs to be able to real-time sense the change in resource requirements of virtual machines and respond quickly to avoid resource waste or performance bottlenecks. This requires the scheduling algorithm to have high efficiency in real-time and flexibility, and be able to dynamically adjust the resource allocation strategy. Summary of the Invention

[0008] The purpose of the present invention is to provide a multi-layer virtual machine scheduling system based on a large language model for the deficiencies of the prior art. This system can provide an efficient, flexible, and adaptive scheduling optimization solution, significantly improving the utilization efficiency of cloud resources.

[0009] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0010] A multi-layer virtual machine scheduling system based on a large language model, the system includes:

[0011] Scheduling scenario feature analysis module: Analyze the historical virtual machine sequence using a large language model, that is, add the virtual machine sequence as input information to the prompt and send it to the large language model. The large language model then automatically extracts the scenario features and classifies them. Its formula is defined as: {s i | i = 1, …, S} = LLM(task_des, S vm ), where S vm represents the virtual machine sequence, including the information of the actually arriving virtual machines, specifically including the resource requirement dimensions and quantities of the virtual machines. task_des represents the task description, including instructions to the large language model to cluster and classify the virtual machine sequence according to the virtual machine resource size. The LLM function is responsible for outputting the processing result of the large language model. {s i | i = 1, …, S} is the classified scenario list;

[0012] Initial heuristic design module: Add the code example and task description to the prompt and send the prompt to the large language model to let the large language model generate an initial workable heuristic scheduling strategy. Its formula is defined as π origin = LLM(task_des, template_cose), thus providing the basic design of the optimization algorithm for the system; where template_code is the example code, task_des is the task description, and the LLM function is responsible for outputting the processing result of the large language model. π origin is the generated initial heuristic scheduling strategy; where the task description covers the scenario description of cloud computing resource scheduling and proposes the requirement for generating the heuristic scheduling strategy;

[0013] Scenario adaptive optimization module: Responsible for processing the virtual machine scheduling sequences from different scenarios {s i | i = 1, …, S}, and independently optimizing the heuristic functions of each scenario using the large language model; add the feature information of each virtual machine scenario, the current optimal heuristic scheduling strategy, and the task description to the prompt respectively, and let the large language model generate the optimized heuristic scheduling strategy π t+1 , and its formula is defined as:

[0014]

[0015] where, is the optimized heuristic scheduling strategy corresponding to each scenario s i , is the current optimal heuristic scheduling strategy corresponding to each scenario s i , task_des is the task description, including instructions to the large language model to perform optimization. It is the feature information of each virtual machine scenario, including the proportion of the five types of virtual machine quantities; this module interacts with the performance evaluation module, and the performance evaluation module conducts performance evaluation on the optimization results and feeds back the evaluation results to the scenario adaptive optimization module; according to the fed-back performance results, the optimization results are sorted, and based on this, the next round of optimization is started to gradually improve the scheduling performance and resource utilization efficiency; among them, the performance result refers to the number of virtual machines that can be successfully placed under fixed resource conditions.

[0016] Performance evaluation module: Use the optimized virtual machine heuristic scheduling strategy to perform simulation placement operations, calculate the number of virtual machines that can be successfully placed under fixed resource conditions, generate performance results, and provide a feedback basis for optimization.

[0017] Multi-scenario strategy management module: Receive different scheduling strategies generated under multiple scenarios provided by the scenario adaptive optimization module; add multiple scheduling strategies {π i | i = 1, …, k}, the current optimal strategy adapter, and the task description to the prompt and send it to the large language model to make it generate a new strategy adapter, and send the strategy list and multi-scenario management strategy to the performance evaluation module; after the performance evaluation module returns the result, and after sorting the strategy adapters, perform the next round of iterative optimization, and its formula is defined as: ξ t+1 = LLM(task_des, ξ * , {π i | i = 1, …, k}), where ξ * is the current optimal strategy adapter, task_des is the task description, including instructing the large language model to optimize ξ * , ξ t+1 is the optimized strategy adapter returned, {π i | i = 1, …, k} is the heuristic scheduling strategy list, and the LLM function is responsible for outputting the processing result of the large language model.

[0018] After the scheduling scenario feature analysis module generates the scenario list, the initial heuristic design module will design corresponding initial heuristic functions according to each scenario as the starting point for subsequent optimization; subsequently, the scenario adaptive optimization module receives the scenario list and its corresponding initial heuristic functions, and in each round of iteration, constructs prompt words and sends them to the large language model to obtain the optimized heuristic functions under each scenario; the optimization results are evaluated by the performance evaluation module for the performance of the heuristic functions, and the next round of optimization iteration is performed according to the evaluation results to gradually improve the function performance.

[0019] The scenario adaptive optimization module provides multiple optional scheduling strategies for the multi-scenario policy management module; after receiving the scheduling policy list, the multi-scenario policy management module designs a policy adapter; the performance evaluation module evaluates the performance of the policy adapter obtained from the large language model in each round of optimization and feeds back the evaluation results to the multi-scenario policy management module.

[0020] Advantageous technical effects of the present invention:

[0021] By introducing a large language model to optimize the virtual machine scheduling heuristic function, the present invention realizes the intelligence and self-adaptability of the multi-layer virtual machine scheduling system. The present invention uses the large language model to automatically generate the initial scheduling strategy, reducing the complexity of manual design. At the same time, the system proposed by the present invention can deeply analyze the characteristics of different scheduling scenarios and dynamically optimize the scheduling strategy according to specific requirements, significantly improving resource utilization and system performance. Through multi-scenario policy management and iterative optimization mechanism, the flexibility and adaptability of the scheduling system are enhanced, ensuring efficient and stable resource allocation in the changing cloud computing environment. The multi-layer virtual machine scheduling system based on the large language model proposed by the present invention effectively improves the intelligence level of virtual machine scheduling and optimizes the efficiency and reliability of cloud resource management. Brief Description of the Drawings

[0022] Figure 1 It is a schematic structural diagram of an embodiment of the present invention. Detailed Embodiments

[0023] To make the technical solutions of the present invention clearer and more definite for those skilled in the art, the present invention will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0024] Embodiment

[0025] A multi-layer virtual machine scheduling system based on a large language model, as Figure 1 shown, includes: a scheduling scenario feature analysis module, an initial heuristic design module, a scenario adaptive optimization module, a performance evaluation module, and a multi-scenario policy management module.

[0026] Among them, the initial heuristic design module is used to generate an initial operable scheduling heuristic function in combination with the large language model, as the basis for the subsequent optimization algorithm of the system; the scheduling scenario feature analysis module is used to analyze the characteristics of historical and current virtual machine scheduling scenarios; the scenario adaptive optimization module communicates with the large language model (LLM) and optimizes the initial heuristic function specifically for different scenarios; the performance evaluation module tests the optimized heuristic function and returns the performance results; the multi-scenario policy management module dynamically screens and updates the available scheduling strategies according to the test results of different scenarios to achieve multi-scenario adaptive management.

[0027] In this embodiment, the scheduling scenario feature analysis module is mainly responsible for deeply analyzing the historical virtual machine sequence and resource requirements, and extracting the key features in different scenarios. The virtual machine sequence information includes the number of virtual machines, the distribution of CPU and memory requirements, and the frequency of arrival time, etc.; the scheduling scenario feature analysis module classifies the extracted scenario information (such as high-load scenarios, burst traffic scenarios, long-tail distribution scenarios, etc.) and outputs it to the scenario adaptive optimization module.

[0028] In this embodiment, the initial heuristic design module, aiming at the requirements of the multi-layer virtual machine scheduling system, uses a large language model to automatically generate a runnable heuristic scheduling function according to the information such as the initial code example and prompt words provided by the user. The user input includes the example code of the existing virtual machine scheduling algorithm, the description of the system's requirements for indicators such as resource utilization or throughput, and the constraint description of multi-dimensional resources (CPU, memory, storage, etc.) in the cloud computing environment. After obtaining the above information, the large language model outputs an initial scheduling heuristic function based on its language understanding and code generation capabilities. The initial heuristic design module integrates this function into the system to form an executable scheduling strategy, laying a foundation for subsequent further optimization.

[0029] In this embodiment, the scenario adaptive optimization module interacts with the large language model to perform targeted optimization on the initial heuristic function or the existing scheduling strategy to adapt to different virtual machine scenarios. The input comes from the scenario information of the scheduling scenario feature analysis module, as well as the existing scheduling strategy or heuristic function; in each iteration, the scenario adaptive optimization module constructs the optimal scheduling heuristic function and scenario description in the current database into a prompt, and sends it to the large language model; the large language model replies with an improved heuristic function, as well as an explanation of the new function and an analysis of its possible advantages and disadvantages; the new function obtained by the module is temporarily stored and passed to the performance evaluation module for testing. If the test result is better than the existing strategy, it is saved in the database.

[0030] In this embodiment, the performance evaluation module conducts simulation or real-scenario tests on the optimized heuristic function to obtain indicators such as the number of virtual machines that can be successfully placed under fixed resource conditions. The evaluation results are presented numerically and fed back to the scenario adaptive optimization module and the multi-scenario policy management module.

[0031] In this embodiment, the main function of the multi-scenario policy management module is to generate a policy adapter through large language model training to screen and manage multiple scheduling policies from the scenario adaptive optimization module, so as to cope with the variability of the cloud environment. According to the results returned by the performance evaluation module, if a certain policy performs excellently in a specific scenario, the module will give priority to recommending or automatically switching to this policy. For the case where multiple scenarios exist in parallel, the module will determine the usage order or priority of the scheduling policies according to the matching degree of scenario characteristics and performance evaluation results to ensure the efficiency and stability of system resource allocation in the dynamically changing cloud computing environment.

[0032] As described above, it is only a further embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention, according to the technical solution and concept of the present invention, makes equivalent substitutions or changes, all belong to the protection scope of the present invention.

Claims

1. A multi-layer virtual machine scheduling system based on a large language model, characterized in that: The system includes: Scheduling scenario feature analysis module: Use the large language model to analyze the historical virtual machine sequence, that is, add the virtual machine sequence as input information to the prompt and send it to the large language model. The large language model then automatically extracts the scenario features and classifies them. The formula is defined as: {s i |i=1,…,S}=LLM(task_des,S vm ), where S vm represents a virtual machine sequence, including the actual virtual machine information, including the resource requirement dimension and quantity of the virtual machine. task_des represents a task description, including instructions to the large language model, requiring it to cluster and classify the virtual machine sequence according to the size of the virtual machine resources. The LLM function is responsible for outputting the processing results of the large language model. i |i=1,…,S} is the classified scene list; Initial heuristic design module: Add code examples and task descriptions to prompts and send the prompts to the large language model, allowing the large language model to generate an initial heuristic scheduling strategy that can be run. The formula is defined as π origin =LLM(task_des,template_code); where template_code is the sample code, task_des is the task description, and the LLM function is responsible for outputting the processing results of the large language model. origin The initial heuristic scheduling strategy is generated; the task description covers the scenario description of cloud computing resource scheduling and proposes the need to generate a heuristic scheduling strategy; Scene adaptive optimization module: responsible for processing i |i=1,…,S}, and use the large language model to independently optimize the heuristic function of each scenario; add the feature information of each virtual machine scenario, the current optimal heuristic scheduling strategy, and the task description to the prompt, and let the large language model generate the optimized heuristic scheduling strategy π t+1 , its formula is defined as: in, For each scene i The corresponding optimized heuristic scheduling strategy, Each scene i The corresponding current optimal heuristic scheduling strategy, task_des is the task description, including the instruction large language model To optimize, It is the characteristic information of each virtual machine scene, including the proportion of the number of five types of virtual machines. This module interacts with the performance evaluation module, which evaluates the optimization results and feeds the evaluation results back to the scene adaptive optimization module. According to the feedback performance results, the optimization results are sorted, and the next round of optimization is started based on this, gradually improving the scheduling performance and resource utilization efficiency. The performance result refers to the number of virtual machines that can be successfully placed under fixed resource conditions. Performance evaluation module: Use the optimized virtual machine heuristic scheduling strategy to simulate placement operations, calculate the number of virtual machines that can be successfully placed under fixed resource conditions, generate performance results, and provide feedback for optimization; Multi-scenario strategy management module: receives different scheduling strategies generated in multiple scenarios provided by the scenario adaptive optimization module; i |i=1,…,k}, the current optimal strategy adapter and task description are added to the prompt and sent to the large language model to generate a new strategy adapter, and the strategy list and multi-scenario management strategy are sent to the performance evaluation module; after the performance evaluation module returns the result, the strategy adapters are sorted and the next round of iterative optimization is carried out. The formula is defined as: t+1 =LLM(task_des,ξ * ,{π i |i=1,…,k}), where ξ * is the current optimal strategy adapter, task_des is the task description, including instructions for the large language model to ξ * To optimize, t+1 is the policy adapter returned after optimization, {π i |i=1,…,k} is a list of heuristic scheduling strategies, and the LLM function is responsible for outputting the processing results of the large language model.

Images