Data center energy optimization method and device based on large language model, medium and product

Abstract

The invention discloses a data center energy optimization method and device based on a large language model, a medium and a product, and relates to the field of data center energy consumption management. The mass data comprises computing resource monitoring data, time sequence feature data, business feature data and environment feature data; carrying out pre-training on the mass data; constructing a domain knowledge graph according to the pre-trained mass data; utilizing an attention mechanism to inject domain knowledge graph knowledge into the large-scale language model, and adopting positive and negative samples with time sequence similarity and a multi-task learning framework to optimize the large-scale language model; carrying out real-time mass data load prediction by utilizing the optimized large language model; and performing multi-objective optimized resource scheduling according to a load prediction result. The data center energy distribution uniformity and the resource utilization rate can be improved, and the energy consumption of the data center can be reduced.

Description

Technical Field

[0001] The present application relates to the field of data center energy consumption management, and in particular to a data center energy optimization method, equipment, medium and product based on a large language model. Background Art

[0002] With the deepening of digital transformation, the energy consumption problem of data centers has become increasingly prominent. Existing data center energy consumption management technologies are mainly based on pre-set static rules and thresholds to manage resources and energy consumption according to fixed logic. Among them, the fixed rule method manages the resources and energy consumption of data centers through pre-defined, non-dynamically adjustable rules. For example: server scheduling shuts down some servers according to fixed time periods (such as nighttime or weekends) to save energy, ignoring real-time load requirements. Assign tasks to specified server groups instead of dynamically adjusting to adapt to the current load distribution. In refrigeration system management, the operating parameters of air conditioners or cooling equipment (such as constant target temperature) are fixed, instead of being dynamically adjusted according to actual load changes. Use timed switch strategies to turn equipment on or off at specific times instead of based on the operating status of real-time data. Resource allocation allocates a fixed number of resources (such as CPU, memory) for each task type, which cannot be flexibly adjusted even if resource requirements change. Simple threshold control methods rely on setting a single threshold for key indicators, and trigger corresponding actions when the monitored value exceeds or falls below the threshold. However, the above methods are only applicable to data centers with smaller scale and relatively stable workloads, and are difficult to cope with complex and changing modern workloads. At present, machine learning methods have made some progress in this field, improving the efficiency of energy consumption management and the intelligent level of resource scheduling in a data-driven way. For example, a multi-objective optimization model of deep learning is used to simultaneously consider energy consumption and temperature control to predict the cooling efficiency of the refrigeration system. However, there are still the following deficiencies in processing long-sequence prediction and feature association of data center energy consumption: 1. Unified representation and efficient processing of heterogeneous data in data centers: The operating environment of data centers is complex, involving multi-dimensional heterogeneous data such as computing resources, business characteristics, time series characteristics, and environmental characteristics. These data are distributed between structured and unstructured data, making them difficult to represent and process in a unified manner.

[0003] 2. The multi-time scale problem of load forecasting: The load changes in data centers show multi-time scale characteristics, including short-term fluctuations (seconds), medium-term changes (hours) and long-term trends (days and above). Traditional forecasting methods are difficult to accurately capture the load characteristics of each time scale at the same time.

[0004] 3. Insufficient model adaptability in data center scenarios: Existing general models cannot fully understand the professional characteristics of data center scenarios, resulting in poor adaptability to load patterns, resource utilization rules and anomaly detection.

[0005] 4. Complexity of multi-objective resource scheduling optimization: Data center resource scheduling involves multi-objective optimization (energy consumption, performance, reliability, etc.), and the scheduling problem is complex, making it difficult for traditional algorithms to achieve global optimization.

[0006] 5. Accuracy issues in data center energy consumption optimization: Traditional energy consumption optimization methods do not fully consider the fine-grained power consumption characteristics of components such as servers, cooling systems, and network equipment, resulting in limited accuracy in energy consumption assessment and optimization results.

[0007] 6. Real-time and adaptability issues of scheduling decisions: The load in data centers changes rapidly and complexly. Traditional scheduling strategies lack real-time and dynamic adaptability, which may lead to uneven resource allocation or excessive energy consumption.

[0008] Therefore, based on the above problems, there is an urgent need to provide a new data center energy optimization method or system. Summary of the invention

[0009] The purpose of this application is to provide a data center energy optimization method, device, medium and product based on a large language model, which can improve the uniformity of data center energy distribution and resource utilization and reduce the energy consumption of the data center.

[0010] To achieve the above objectives, this application provides the following solutions: In a first aspect, the present application provides a data center energy optimization method based on a large language model, and the data center energy optimization method based on a large language model includes: Acquire massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is the load fluctuation and periodic law captured by the multi-scale sampling strategy; the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; Pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction, and load similarity judgment; Build domain knowledge graph based on massive pre-trained data; The attention mechanism is used to inject domain knowledge graph knowledge into large language models, and the positive and negative samples of temporal similarity and the multi-task learning framework are used to optimize large language models. Use the optimized large-scale language model to perform real-time load forecasting of massive data; the load forecast results include: short-term load forecasting, medium-term load forecasting and long-term load forecasting; short-term load forecasting is used to predict load trends at the minute to hour level; medium-term load forecasting is used to predict load trends at the day to week level; long-term load forecasting is used to predict monthly load trends; Perform multi-objective optimized resource scheduling based on load forecast results.

[0011] Optionally, the multi-task learning framework is constructed based on a deep learning framework; the deep learning framework includes: PyTorch and TensorFlow.

[0012] Optionally, the multi-objective optimized resource scheduling according to the load prediction result also includes: Fusion of load forecast results; The weights of different time scales are dynamically adjusted according to the fusion results, and the final load forecast results are output.

[0013] Optionally, multi-objective optimized resource scheduling is performed according to the load prediction result, specifically including: According to the final load prediction results, a random forest regression model is used to build an energy consumption model; the energy consumption model is used to dynamically determine the power consumption of servers, cooling systems, and network equipment; Determine evaluation performance indicators and reliability indicators based on the energy consumption model; the evaluation performance indicators include: response time, throughput and resource utilization; the reliability indicators include: fault tolerance and fault recovery capability; According to the Pareto optimal principle and evaluation performance indicators and reliability indicators, the optimization scheduling objective function is constructed; According to the optimization scheduling objective function, a hierarchical reinforcement learning framework and a policy network are used to obtain a resource scheduling solution; Adjust server load distribution, resource scheduling and cooling system according to the resource scheduling plan.

[0014] Optionally, the hierarchical reinforcement learning framework is built on the Options Framework algorithm.

[0015] Optionally, the optimization scheduling objective function is: ; Using the formula Determine task allocation constraints; Using the formula Identify time constraints; Using the formula Determine reliability constraints; in, are weight coefficients used to reflect the importance of energy consumption, performance and reliability targets, respectively. is the static power consumption of server i, is the dynamic power consumption coefficient, For server load, is the average load of all servers in the system, is the cooling system energy consumption associated with server i, is the response time of task j, which is related to the scheduling scheme x, is the total number of tasks, Is task j assigned to server i? 1 means assigned, 0 means not assigned. is the capacity limit of server i, is the maximum allowed response time of the task, is the minimum reliability level required by the system, is the reliability level, a function of the scheduling scheme x, which measures the reliability of the system under a given task allocation and resource scheduling scheme. N is the total number of servers.

[0016] In a second aspect, the present application provides a data center energy optimization device based on a large language model, and the data center energy optimization device based on a large language model includes: The data acquisition module is used to acquire massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is the load fluctuation and periodic law captured by the multi-scale sampling strategy; the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; A data pre-training module, used for pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction and load similarity judgment; The domain knowledge graph construction module is used to construct the domain knowledge graph based on the massive pre-trained data; Large language model optimization module, which uses the attention mechanism to inject domain knowledge graph knowledge into large language models, and uses positive and negative samples of temporal similarity and a multi-task learning framework to optimize large language models; The load prediction module is used to use the optimized large-scale language model to perform real-time load prediction of massive data; the load prediction results include: short-term load prediction, medium-term load prediction and long-term load prediction; short-term load prediction is used to predict load trends at the minute to hour level; medium-term load prediction is used to predict load trends at the day to week level; long-term load prediction is used to predict monthly load trends; The resource scheduling module is used to perform multi-objective optimized resource scheduling based on load prediction results.

[0017] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the data center energy optimization method based on a large language model.

[0018] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data center energy optimization method based on a large language model.

[0019] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the data center energy optimization method based on a large language model.

[0020] According to the specific embodiments provided in this application, this application has the following technical effects: The present application provides a data center energy optimization method, device, medium and product based on a large language model, which pre-trains the massive data to solve the problem that the data is distributed between structured and unstructured and is difficult to uniformly represent and process; uses an attention mechanism to inject domain knowledge graph knowledge into the large language model, and uses positive and negative samples of temporal similarity and a multi-task learning framework to optimize the large language model; fully understands the professional characteristics of the data center scenario and improves the adaptability to load patterns, resource utilization rules and anomaly detection; uses the optimized large language model to perform real-time load prediction of massive data; performs multi-objective optimized resource scheduling based on the load prediction results; fully considers the fine-grained power consumption characteristics of components such as servers, refrigeration systems, and network equipment to improve the accuracy of energy consumption evaluation and optimization results; and performs multi-objective optimized resource scheduling based on the prediction of real-time massive data, which can improve the uniformity of energy distribution and resource utilization in the data center and reduce the energy consumption of the data center. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 This is a flow chart of a data center energy optimization method based on a large language model in one embodiment of the present application. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0024] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0025] In an exemplary embodiment, Figure 1 As shown, a data center energy optimization method based on a large language model is provided, and the method includes the following S101 to S106. Among them: S101, obtaining massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is load fluctuation and periodic law captured by multi-scale sampling strategy (seconds, minutes, hours, days); the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; Key indicators in computing resource monitoring data use high-frequency sampling (1 second), and non-key indicators use low-frequency sampling (minutes). Key indicators include: CPU usage, memory access mode, network bandwidth utilization, storage IO performance, etc. After obtaining the computing resource monitoring data, the collection frequency is adjusted dynamically according to the importance. Importance is based on the current operating status of the data center, predicting the current load trend and performance bottleneck, and dynamically adjusting the indicator priority according to the prediction results: 1. When the system load is high, the CPU and memory-related indicators have a higher priority, and the sampling frequency of CPU usage and memory access patterns is increased. 2. When the task execution involves a large amount of data transmission, the priority of network and storage-related indicators is increased, and the sampling frequency is increased.

[0026] Among them, periodicity refers to the specific periodic change pattern of load over time, usually for a longer time range. Load fluctuation refers to the instantaneous change of resource usage during the operation of the load system, mainly for dynamic changes in a short time range.

[0027] Business feature data converts unstructured business information (task type, resource requirements, etc.) into structured data through a standardized task description framework. The standardized task description framework abstracts, categorizes, and quantifies unstructured task information to achieve efficient and unified data representation; The core elements of the standardized task description framework are as follows: 1. Task type, by workload: Online tasks (real-time processing), offline tasks (batch processing); and computing characteristics: computing-intensive tasks (high CPU usage), IO-intensive tasks (frequent disk or network operations), standardized classification; 2. Resource requirements: The distance types are mainly: CPU: Number of cores and frequency required.

[0028] Memory: Required capacity and access speed.

[0029] Network: Bandwidth requirements and latency tolerance.

[0030] Storage: IOPS requirements and storage capacity.

[0031] 3. Service Level: According to the priority of the task and the requirements of service quality, it is divided into three levels: critical, important and ordinary; 4. Time constraints: According to the time requirements of the task, quantify the shortest completion time and the latest completion time.

[0032] 5. Runtime dependencies: According to the description of the software, hardware and other tasks that the task depends on, it is divided into hardware dependency (GPU model, storage type), software dependency (operating environment, version), and task dependency (task ID that must be completed first) 6. Environmental requirements: The unique requirements of the mission's operating environment, including temperature ranges and power requirements.

[0033] S102, pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction, and load similarity judgment; Among them, data compression algorithms include: lossless compression, lossy compression, time series data and streaming data compression; real-time compression of massive collected data to reduce storage and transmission overhead.

[0034] Aggregate time series data (mainly including: CPU temperature, CPU usage, bandwidth utilization, memory usage and storage read and write speed, room temperature and humidity, cooling efficiency, electricity price and weather temperature and humidity, task load data, etc.) through the sliding window mechanism, and intelligently downsample historical data to retain key information points (data processing based on the sliding window mechanism, dynamically adjust the window size according to the data change rate or mutation frequency, reduce the window when the change is drastic, and increase the window when it is stable); S103, constructing a domain knowledge graph based on the massive pre-trained data; S104, using the attention mechanism to inject domain knowledge graph knowledge into the large language model, and using positive and negative samples of temporal similarity and a multi-task learning framework to optimize the large language model; the deep learning framework includes: PyTorch and TensorFlow.

[0035] Among them, the model is optimized by using positive and negative sample pairs with time series similarity to enhance its ability to identify load patterns and abnormal patterns; the multi-task learning framework is used to achieve knowledge sharing between tasks and improve prediction accuracy; new data is continuously collected to optimize large language models to adapt to dynamically changing loads and environments In the optimized large language model, the local attention mechanism is used to capture load changes in real time, and the attention window is dynamically adjusted to cope with high-frequency fluctuations; residual connections and position encoding are combined to enhance the ability to capture short-term loads; through Fourier analysis, the time series data is converted from the time domain to the frequency domain, so as to identify potential periodic patterns in the data (periodic patterns mainly include intraday cycles, weekly cycles, task scheduling cycles, electricity price cycles and environmental cycles) and extract features (periodic features, trend features), and separate trend and periodic components. Dynamically adjust the weight of the periodic pattern to achieve accurate prediction of the medium-term load; use trend decomposition methods (such as additive / multiplicative decomposition, HP filter, STL, wavelet decomposition, etc.) to model long-term trends, seasonal changes and random fluctuations, and integrate multi-component prediction results to improve long-term prediction accuracy.

[0036] S105, using the optimized large language model to perform real-time load prediction of massive data; the load prediction results include: short-term load prediction, medium-term load prediction and long-term load prediction; the short-term load prediction is used to predict the load trend at the minute to hour level; the medium-term load prediction is used to predict the load trend at the day to week level; the long-term load prediction is used to predict the monthly load trend; In short-term load forecasting, the local attention mechanism is used to dynamically adjust the attention window to adapt to high-frequency load changes. Multi-head attention and residual connection techniques are applied to capture fine-grained short-term changes; Use Fourier analysis to detect periodic patterns and extract periodic features in medium-term load forecasting; separate trend and periodic components and use adaptive fusion strategies to weight forecasts at different time scales; In long-term load forecasting, the trend is decomposed into long-term trend, seasonal pattern and random fluctuation, and modeled separately; the accuracy of long-term forecast is improved by fusing multi-component forecast results; S106, performing multi-objective optimized resource scheduling according to the load prediction result.

[0037] Before S106, it also included: Fusion of load forecast results; The weights of different time scales are dynamically adjusted according to the fusion results, and the final load forecast results are output.

[0038] S106 specifically includes: According to the final load prediction results, a random forest regression model is used to build an energy consumption model; the energy consumption model is used to dynamically determine the power consumption of servers, cooling systems, and network equipment; Determine evaluation performance indicators and reliability indicators based on the energy consumption model; the evaluation performance indicators include: response time, throughput and resource utilization; the reliability indicators include: fault tolerance and fault recovery capability; According to the Pareto optimal principle and evaluation performance indicators and reliability indicators, the optimization scheduling objective function is constructed; According to the optimization scheduling objective function, the hierarchical reinforcement learning framework and policy network are used to obtain the resource scheduling solution; the hierarchical reinforcement learning framework is built based on the Options Framework algorithm. The hierarchical reinforcement learning framework is used to decompose complex scheduling problems into multiple sub-problems (such as resource allocation, load balancing, and cooling system optimization); the policy network generates the optimal scheduling solution and dynamically adjusts server resources and environmental configuration; Policy network is a model structure in deep learning and reinforcement learning. It is mainly used to learn a strategy. The policy network can choose the appropriate action according to the current state. The core task of the policy network is to directly output the probability distribution of an action for each possible state, or directly output a specific action. In reinforcement learning, the policy network can improve its decision-making effect in the environment through the optimization process. The policy network can dynamically generate the following scheduling schemes: (1) Task allocation: The tasks are assigned to suitable servers to minimize the response time and energy consumption to determine the objective function.

[0039] Distribute tasks to servers with lighter loads.

[0040] Optimize the allocation strategy of high-priority tasks to ensure that tasks are completed in the shortest possible time.

[0041] (2) Load balancing: Balance the load between servers to avoid overloading a single point.

[0042] Select the source and destination servers for the migration.

[0043] Dynamically adjust the operating status of servers (such as enabling / disabling idle servers).

[0044] (3) Refrigeration system optimization: Dynamically adjust cooling system parameters to minimize energy consumption.

[0045] Adjust the coolant temperature.

[0046] Optimize fan speed or air flow distribution.

[0047] (4) Anomaly detection and processing: Identify and resolve abnormal conditions in the system (such as high temperature or degraded performance).

[0048] Trigger the migration plan for the failed server.

[0049] Adjust the priority of related servers to ensure system stability.

[0050] Adjust server load distribution, resource scheduling and cooling system according to the resource scheduling plan.

[0051] The optimization scheduling objective function is: ; Using the formula Determine task allocation constraints; Using the formula Identify time constraints; Using the formula Determine reliability constraints; in, are weight coefficients used to reflect the importance of energy consumption, performance and reliability targets, respectively. is the static power consumption of server i, is the dynamic power consumption coefficient, For server load, is the average load of all servers in the system, is the cooling system energy consumption associated with server i, is the response time of task j, which is related to the scheduling scheme x, is the total number of tasks, Is task j assigned to server i? 1 means assigned, 0 means not assigned. is the capacity limit of server i, is the maximum allowed response time of the task, is the minimum reliability level required by the system, is the reliability level, a function of the scheduling scheme x, which measures the reliability of the system under a given task allocation and resource scheduling scheme. N is the total number of servers.

[0052] S106 and later also include: Monitor system response in real time to verify scheduling effects. Collect execution results, including energy consumption, performance, and reliability data, and analyze the effectiveness of scheduling strategies. Use feedback data to update energy consumption models and reinforcement learning frameworks to continuously optimize scheduling strategies. Dynamically update models and strategies during daily operations to keep the system efficiently adaptable to environmental changes.

[0053] Based on the same inventive concept, the embodiment of the present application also provides a data center energy optimization device based on a large language model for implementing the data center energy optimization method based on a large language model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more data center energy optimization device embodiments based on a large language model provided below can refer to the limitations of the data center energy optimization method based on a large language model above, and will not be repeated here.

[0054] In an exemplary embodiment, a data center energy optimization device based on a large language model is provided, comprising: The data acquisition module is used to acquire massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is the load fluctuation and periodic law captured by the multi-scale sampling strategy; the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; A data pre-training module, used for pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction and load similarity judgment; The domain knowledge graph construction module is used to construct the domain knowledge graph based on the massive pre-trained data; Large language model optimization module, which uses the attention mechanism to inject domain knowledge graph knowledge into large language models, and uses positive and negative samples of temporal similarity and a multi-task learning framework to optimize large language models; The load prediction module is used to use the optimized large-scale language model to perform real-time load prediction of massive data; the load prediction results include: short-term load prediction, medium-term load prediction and long-term load prediction; short-term load prediction is used to predict load trends at the minute to hour level; medium-term load prediction is used to predict load trends at the day to week level; long-term load prediction is used to predict monthly load trends; The resource scheduling module is used to perform multi-objective optimized resource scheduling based on load prediction results.

[0055] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a data center energy optimization method based on a large language model is implemented.

[0056] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0057] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0058] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0059] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0060] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0061] In this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0062] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A data center energy optimization method based on a large language model, characterized in that: The data center energy optimization method based on the large language model includes: Acquire massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is the load fluctuation and periodic law captured by the multi-scale sampling strategy; the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; Pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction, and load similarity judgment; Build domain knowledge graph based on massive pre-trained data; The attention mechanism is used to inject domain knowledge graph knowledge into large language models, and the positive and negative samples of temporal similarity and the multi-task learning framework are used to optimize large language models. Use the optimized large-scale language model to perform real-time load forecasting of massive data; the load forecast results include: short-term load forecasting, medium-term load forecasting and long-term load forecasting; short-term load forecasting is used to predict load trends at the minute to hour level; medium-term load forecasting is used to predict load trends at the day to week level; long-term load forecasting is used to predict monthly load trends; Perform multi-objective optimized resource scheduling based on load forecast results.

2. The data center energy optimization method based on a large language model according to claim 1, characterized in that: The multi-task learning framework is built based on a deep learning framework; the deep learning framework includes: PyTorch and TensorFlow.

3. The data center energy optimization method based on a large language model according to claim 1, characterized in that: The resource scheduling for multi-objective optimization according to the load prediction result also includes: Fusion of load forecast results; The weights of different time scales are dynamically adjusted according to the fusion results, and the final load forecast results are output.

4. The data center energy optimization method based on a large language model according to claim 3 is characterized in that: Multi-objective optimization resource scheduling based on load forecast results, including: According to the final load prediction results, a random forest regression model is used to build an energy consumption model; the energy consumption model is used to dynamically determine the power consumption of servers, cooling systems, and network equipment; Determine evaluation performance indicators and reliability indicators based on the energy consumption model; the evaluation performance indicators include: response time, throughput and resource utilization; the reliability indicators include: fault tolerance and fault recovery capability; According to the Pareto optimal principle and evaluation performance indicators and reliability indicators, the optimization scheduling objective function is constructed; According to the optimization scheduling objective function, a hierarchical reinforcement learning framework and a policy network are used to obtain a resource scheduling solution; Adjust server load distribution, resource scheduling and cooling system according to the resource scheduling plan.

5. The data center energy optimization method based on a large language model according to claim 4, characterized in that: The hierarchical reinforcement learning framework is built on the Options Framework algorithm.

6. The data center energy optimization method based on a large language model according to claim 4, characterized in that: The optimization scheduling objective function is: ； Using the formula Determine task allocation constraints; Using the formula Identify time constraints; Using the formula Determine reliability constraints; in, are weight coefficients used to reflect the importance of energy consumption, performance and reliability targets, respectively. is the static power consumption of server i, is the dynamic power consumption coefficient, For server load, is the average load of all servers in the system, is the cooling system energy consumption associated with server i, is the response time of task j, which is related to the scheduling scheme x, is the total number of tasks, Is task j assigned to server i? 1 means assigned, 0 means not assigned. is the capacity limit of server i, is the maximum allowed response time of the task, is the minimum reliability level required by the system, is the reliability level, a function of the scheduling scheme x, which measures the reliability of the system under a given task allocation and resource scheduling scheme. N is the total number of servers.

7. A data center energy optimization device based on a large language model, characterized in that: The data center energy optimization device based on the large language model includes: The data acquisition module is used to acquire massive data; the massive data includes: computing resource monitoring data, time series feature data, business feature data and environmental feature data; the computing resource monitoring data includes: CPU usage, memory access mode, network bandwidth utilization and storage IO performance; the time series feature data is the load fluctuation and periodic law captured by the multi-scale sampling strategy; the business feature data includes: task type, resource demand, service level and time constraint; the environmental feature data includes: integrated external environment data and computer room operation data; the external environment data includes: electricity price, weather information and holidays; the computer room operation data includes: temperature and humidity and cooling efficiency; A data pre-training module, used for pre-training the massive data; the pre-training includes: data compression, data aggregation, load sequence completion, anomaly detection, resource utilization prediction and load similarity judgment; The domain knowledge graph construction module is used to construct the domain knowledge graph based on the massive pre-trained data; Large language model optimization module, which uses the attention mechanism to inject domain knowledge graph knowledge into large language models, and uses positive and negative samples of temporal similarity and a multi-task learning framework to optimize large language models; The load prediction module is used to use the optimized large-scale language model to perform real-time load prediction of massive data; the load prediction results include: short-term load prediction, medium-term load prediction and long-term load prediction; short-term load prediction is used to predict load trends at the minute to hour level; medium-term load prediction is used to predict load trends at the day to week level; long-term load prediction is used to predict monthly load trends; The resource scheduling module is used to perform multi-objective optimized resource scheduling based on load prediction results.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the data center energy optimization method based on a large language model as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the data center energy optimization method based on a large language model described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the data center energy optimization method based on a large language model described in any one of claims 1 to 6 is implemented.

Images