Cloud computing facility control method, device, equipment, storage medium and program product

Abstract

This disclosure relates to a method, apparatus, device, storage medium, and program product for controlling cloud computing facilities. The method includes: acquiring multimodal operational data of each node in the cloud computing facility based on various heterogeneous types of sensors; dynamically allocating the weights of features of each modality based on a long short-term memory network ensemble model combined with an attention mechanism to predict load forecast values ​​for multiple future time steps; determining the load fluctuation amplitude and load change trend based on the load forecast values, dynamically adjusting low-load trigger thresholds and high-load trigger thresholds, acquiring real-time load data of each node, and determining the load status of each node based on the low-load trigger thresholds, high-load trigger thresholds, and real-time load data; performing energy-saving operations on nodes in a low-load state and performing capacity expansion operations on nodes in a high-load state. This method predicts load trends in advance, reduces additional energy consumption caused by frequent power-on and power-off cycles, and improves energy efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of cloud computing technology, and in particular to a cloud computing facility control method, apparatus, equipment, storage medium, and program product. Background Technology

[0002] This section is intended to provide background or context for embodiments of this disclosure. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] As cloud computing continues to expand, the energy consumption of data centers is becoming increasingly prominent. Existing cloud computing energy-saving technologies have mainly gone through several stages of development: from early hardware-level optimization (such as low-power servers), to scheduling strategies based on static thresholds (such as setting fixed CPU utilization thresholds to trigger node sleep and wake-up), and then to strategies that introduce dynamic thresholds and virtual machine consolidation.

[0004] However, most cloud computing energy-saving solutions in related technologies rely on real-time load to trigger energy-saving operations, which is a reactive approach. This lack of foresight can easily lead to frequent node power-on and power-off when faced with sudden traffic surges, not only failing to save energy effectively but also increasing additional energy consumption and hardware wear and tear. Summary of the Invention

[0005] In order to at least partially solve one of the technical problems in the related technologies, this disclosure provides a cloud computing facility control method, apparatus, equipment, storage medium and program product.

[0006] In view of the above objectives, the first aspect of the exemplary embodiments of this disclosure provides a cloud computing facility control method, the method comprising: Multimodal operation data of each node of the cloud computing facility is acquired based on various heterogeneous types of sensors, and the multimodal operation data is preprocessed to obtain a standardized historical feature vector. The standardized historical feature vectors are input into a pre-trained long short-term memory network ensemble model. The pre-trained long short-term memory network ensemble model, combined with an attention mechanism, dynamically allocates the weights of features of each modality and predicts the load prediction values ​​for multiple future time steps. Based on the load forecast value, determine the load fluctuation range and load change trend, dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation range and load change trend, obtain real-time load data of each node, and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold and the real-time load data. For any of the nodes, in response to the node's load state being low load, an energy-saving operation is performed on the node in the low load state; in response to the node's load state being high load, a capacity expansion operation is performed on the node in the high load state.

[0007] In some exemplary embodiments, the step of inputting the standardized historical feature vector into a pre-trained long short-term memory network ensemble model, and dynamically allocating the weights of features of each modality through the pre-trained long short-term memory network ensemble model combined with an attention mechanism to predict the load prediction values ​​for multiple future time steps includes: The standardized historical feature vector is input into the pre-trained long short-term memory network ensemble model, and the temporal features are extracted through the pre-trained long short-term memory network ensemble model to generate a hidden state sequence. The hidden state sequence is input into the attention mechanism. Based on the attention mechanism, the correlation scores between each time step and each modality feature and the current prediction task are calculated. After normalization, the attention weights of each modality feature are obtained. The attention weights are summed with their corresponding hidden states to obtain the context vector. The context vector is input into the fully connected layer, which outputs the load prediction values ​​for multiple future time steps.

[0008] In some exemplary embodiments, the steps of determining the load fluctuation amplitude and load change trend based on the load forecast value, dynamically adjusting the low load trigger threshold and the high load trigger threshold based on the load fluctuation amplitude and the load change trend, obtaining real-time load data for each node, and determining the load status of each node based on the low load trigger threshold, the high load trigger threshold, and the real-time load data include: A load forecast value sequence is constructed based on the load forecast value, and statistical feature calculations are performed based on the load forecast value sequence to determine the load fluctuation amplitude and the load change trend. Using preset mapping rules, the low load trigger threshold and the high load trigger threshold are dynamically adjusted based on the load fluctuation amplitude and the load change trend; The real-time load data is compared with the low load trigger threshold and the high load trigger threshold to determine the load state of each node, which includes a low load state and a high load state.

[0009] In some exemplary embodiments, the step of performing energy-saving operations on nodes in the low-load state in response to the node's load state being a low-load state includes: Cache the image of the operating system and application environment of the node in the low load state locally; A soft shutdown command is sent to the node in the low load state through the intelligent platform management interface, causing the node in the low load state to enter a low-power standby state that retains power only from the baseboard management controller.

[0010] In some exemplary embodiments, the scaling operation performed on the node in the high-load state in response to the node's load state includes: The hibernation nodes are sorted based on hardware performance indicators, and the hibernation node with the best hardware performance is selected as the wake-up target node. The system sends a power-on command to the wake-up target node through the intelligent platform management interface and configures the pre-boot execution environment to enable the wake-up target node to load the locally cached image, skip the hardware self-test process, complete the power-on, and automatically connect to the resource pool. Migrate the application or virtual machine on the node that is in a high-load state to the wake-up target node.

[0011] In some exemplary embodiments, the multimodal operational data includes: CPU utilization, memory usage, network bandwidth utilization, disk I / O activity, and device surface temperature.

[0012] Based on the same inventive concept, a second aspect of the exemplary embodiments of this disclosure provides a cloud computing facility control device, the device comprising: The multimodal operation data determination module is configured to acquire multimodal operation data of each node of the cloud computing facility based on multiple heterogeneous types of sensors, preprocess the multimodal operation data, and obtain a standardized historical feature vector. The load prediction value determination module is configured to input the standardized historical feature vector into a pre-trained long short-term memory network ensemble model, and dynamically allocate the weights of the features of each modality through the pre-trained long short-term memory network ensemble model combined with an attention mechanism to predict the load prediction values ​​for multiple future time steps. The load status determination module is configured to determine the load fluctuation amplitude and load change trend based on the load prediction value, dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation amplitude and the load change trend, obtain real-time load data of each node, and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold and the real-time load data. The control execution module is configured to, for any of the nodes, perform energy-saving operations on nodes in a low-load state in response to the node's load state being low-load, and perform capacity expansion operations on nodes in a high-load state in response to the node's load state being high-load.

[0013] Based on the same inventive concept, a third aspect of the exemplary embodiments of this disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method as described in the first aspect.

[0014] Based on the same inventive concept, a fourth aspect of the exemplary embodiments of this disclosure provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0015] Based on the same inventive concept, a fifth aspect of the exemplary embodiments of this disclosure provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method as described in the first aspect.

[0016] The technical solution provided in this disclosure has the following advantages compared with the prior art: By predicting load trends in advance, the extra energy consumption caused by frequent power-on and power-off cycles is reduced, thus improving energy efficiency. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a cloud computing facility control method provided for an exemplary embodiment of this disclosure; Figure 2 A schematic diagram of a cloud computing facility control system provided as an exemplary embodiment of the present disclosure; Figure 3 A schematic diagram of a cloud computing facility control device provided for an exemplary embodiment of the present disclosure; Figure 4 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this disclosure. Detailed Implementation

[0020] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0021] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0022] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0023] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0024] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0025] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0026] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0027] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0028] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0029] As described in the background section, the inventors of this disclosure have found that most cloud computing energy-saving solutions in related technologies rely on real-time load to trigger energy-saving operations, which is a reactive measure. This lack of foresight can easily lead to frequent node power-on and power-off when faced with sudden traffic surges, not only failing to effectively save energy but also increasing additional energy consumption and hardware wear.

[0030] To address the aforementioned problems, this disclosure provides a cloud computing facility control method, which will be described below with reference to specific embodiments.

[0031] Figure 1 This is a flowchart illustrating a cloud computing facility control method provided in an embodiment of this disclosure. The method can be executed by a cloud computing facility control device, which can be implemented using software and / or hardware, and is generally integrated into an electronic device. Figure 1 As shown, the method includes the following steps: Step 110: Acquire multimodal operation data of each node of the cloud computing facility based on various heterogeneous types of sensors, preprocess the multimodal operation data to obtain standardized historical feature vectors.

[0032] In some exemplary embodiments, the multimodal operational data includes: CPU utilization, memory usage, network bandwidth utilization, disk I / O activity, and device surface temperature.

[0033] Among them, various heterogeneous types of sensors and performance counters are deployed on each computing node of the cloud computing infrastructure to form a multimodal perception layer.

[0034] The various heterogeneous sensor types include at least: a performance counter for acquiring CPU (Central Processing Unit) utilization, a performance monitoring unit for acquiring memory usage, a network interface controller for acquiring network bandwidth utilization, a disk controller for acquiring disk I / O (input / output) activity, and a temperature sensor (such as a thermistor or thermocouple) for acquiring device surface temperature. These sensors are heterogeneous in physical characteristics, sampling frequency, and output data type; for example, temperature sensors output analog or digital signals, while performance counters output discrete percentage values. Through these various heterogeneous sensor types, this disclosure can acquire multi-dimensional operational data in real time, encompassing computing resources, memory resources, network resources, storage resources, and the physical environment—that is, multimodal operational data.

[0035] In this paper, considering that outliers may be present in the multimodal operating data due to occasional sensor failures or network interruptions, this disclosure performs outlier removal and normalization preprocessing on the multimodal operating data to eliminate the impact of outliers on the accuracy of subsequent predictions.

[0036] As an example, the 3σ principle (Raida criterion) is used to remove outliers from the data in each dimension.

[0037] As an example, after outlier removal, in order to eliminate the influence of different units on model training, this disclosure performs Min-Max (minimum-maximum) normalization on the cleaned data, mapping the values ​​of each dimension to the interval [0, 1].

[0038] After the above preprocessing, this disclosure combines the normalized data from 60 consecutive time points (60 time steps, corresponding to the multimodal operation data of the past 60 seconds) into a standardized historical feature vector.

[0039] Cloud computing energy-saving solutions in related technologies typically only collect IT (information technology) resource indicators such as CPU utilization and memory usage, ignoring key dimensions that also affect node load and energy consumption, such as network bandwidth utilization, disk I / O activity, and device surface temperature. This single-dimensional perception method cannot comprehensively reflect the true operating status of nodes, leading to biased load judgments. It may easily misjudge nodes that are actually under high load as low load and shut them down incorrectly, or cause hardware failure risks due to failure to detect abnormal temperatures.

[0040] To address the aforementioned issues, this disclosure employs a multimodal perception index combination scheme. By deploying a distributed acquisition agent, it collaboratively collects operational data across five dimensions: CPU utilization, memory usage, network bandwidth utilization, disk I / O activity, and device surface temperature. The collected data undergoes outlier removal and normalization preprocessing to generate standardized multimodal historical feature vectors.

[0041] Thus, this disclosure enables a comprehensive perception of the operational status of cloud computing infrastructure nodes, significantly improves the accuracy of load assessment, avoids misjudgments caused by a single perception dimension, and provides data support for hardware health management by introducing temperature indicators.

[0042] Step 120: Input the standardized historical feature vector into the pre-trained long short-term memory network ensemble model. Through the pre-trained long short-term memory network ensemble model combined with the attention mechanism, dynamically allocate the weights of the features of each modality and predict the load prediction values ​​for multiple future time steps.

[0043] In some exemplary embodiments, the step of inputting the standardized historical feature vector into a pre-trained long short-term memory network ensemble model, and dynamically allocating the weights of features of each modality through the pre-trained long short-term memory network ensemble model combined with an attention mechanism to predict the load prediction values ​​for multiple future time steps includes: The standardized historical feature vector is input into the pre-trained long short-term memory network ensemble model, and the temporal features are extracted through the pre-trained long short-term memory network ensemble model to generate a hidden state sequence. The hidden state sequence is input into the attention mechanism. Based on the attention mechanism, the correlation scores between each time step and each modality feature and the current prediction task are calculated. After normalization, the attention weights of each modality feature are obtained. The attention weights are summed with their corresponding hidden states to obtain the context vector. The context vector is input into the fully connected layer, which outputs the load prediction values ​​for multiple future time steps.

[0044] The standardized historical feature vector (60×5 dimensions, i.e., 60 historical time steps and 5 modal features) is input into a pre-trained LSTM (Long Short-Term Memory) ensemble model. This LSTM ensemble model consists of three stacked LSTM layers with 64, 128, and 64 neurons respectively, used to extract deep temporal dependencies from multimodal time-series data. For the input historical feature vector, the three LSTM layers process the input sequence step by step through a gating mechanism, generating a corresponding hidden state at each time step. After processing all 60 time steps, a hidden state sequence is obtained, which encodes the multimodal fusion information of each time point within the historical 60 seconds.

[0045] To dynamically highlight the importance of different modal features to the load prediction task at different time steps, this invention introduces an attention mechanism module after the LSTM layer. The attention mechanism module receives the hidden state sequence output by the LSTM and dynamically allocates weights by calculating the correlation scores between each time step and each modal feature and the current prediction task. Taking an additive attention mechanism as an example, for the hidden state at each time step, it is first decomposed into five sub-vectors, corresponding to the encoded representations of five modalities: CPU, memory, network, disk, and temperature. For each modality, the attention score for that modality at that time step is calculated. Subsequently, the scores of the same modality at different time steps are normalized using an activation function (e.g., softmax (logistic regression) function) to obtain the time-step attention weight for that modality. The attention weight reflects the importance of modal information at historical moments when predicting the current load. Then, the hidden states of the same modality at different time steps are weighted and summed with their corresponding attention weights to obtain the context vector for that modality. The context vector aggregates the temporal information of the modalities in the historical window and achieves focus on key time steps through the attention mechanism. Finally, the context vectors of all five modalities are concatenated to form a comprehensive feature representation that integrates multimodal time-series information. This representation is then input into a fully connected layer to output the load prediction values ​​for the next 10 time steps, with each time step corresponding to 30 seconds, thus predicting the load change trend for the next 5 minutes.

[0046] It should be noted that the specific implementation of the attention mechanism is not limited to additive attention. Those skilled in the art can also use variations such as dot product attention, scaled dot product attention, or multi-head attention. As long as dynamic weighting of features at each time step and each modality can be achieved, they all fall within the scope of protection of this disclosure. In addition, the number of layers, number of neurons, and input / output time steps of the LSTM ensemble model are exemplary configurations and can be adjusted according to specific scenarios in actual applications.

[0047] Cloud computing energy-saving solutions in related technologies typically rely on real-time load triggering of energy-saving operations, which is a reactive management model. This model lacks forward-looking prediction when facing sudden traffic surges, easily leading to frequent node power-on and power-off cycles. This not only fails to effectively save energy but also consumes more energy due to the additional power-on and power-off operations, while also increasing hardware wear and tear.

[0048] To address the aforementioned issues, this disclosure employs a load prediction scheme that combines an attention mechanism with an LSTM ensemble model. The multimodal historical feature vector is input into a three-layer LSTM network to extract temporal features. The weights of each time step and each modal feature are dynamically calculated through the attention mechanism, highlighting the contribution of key modalities under the current load mode, and outputting load prediction values ​​for multiple future time steps.

[0049] Therefore, this disclosure enables forward-looking prediction of load trends, allowing for early detection of load changes, avoiding frequent power on / off operations caused by sudden traffic surges, reducing additional energy consumption, improving overall energy efficiency, and providing a reliable decision-making basis for dynamic threshold adjustment.

[0050] Step 130: Determine the load fluctuation amplitude and load change trend based on the load prediction value; dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation amplitude and the load change trend; obtain real-time load data of each node; and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold, and the real-time load data.

[0051] In some exemplary embodiments, the steps of determining the load fluctuation amplitude and load change trend based on the load forecast value, dynamically adjusting the low load trigger threshold and the high load trigger threshold based on the load fluctuation amplitude and the load change trend, obtaining real-time load data for each node, and determining the load status of each node based on the low load trigger threshold, the high load trigger threshold, and the real-time load data include: A load forecast value sequence is constructed based on the load forecast value, and statistical feature calculations are performed based on the load forecast value sequence to determine the load fluctuation amplitude and the load change trend. Using preset mapping rules, the low load trigger threshold and the high load trigger threshold are dynamically adjusted based on the load fluctuation amplitude and the load change trend; The real-time load data is compared with the low load trigger threshold and the high load trigger threshold to determine the load state of each node, which includes a low load state and a high load state.

[0052] Specifically, the load fluctuation amplitude and load change trend are determined based on load forecast values ​​at multiple future time steps. The load forecast values ​​at multiple future time steps constitute a load forecast value sequence, and the load fluctuation amplitude is determined by calculating the statistical characteristics of this load forecast value sequence. As an example, this disclosure uses standard deviation as a quantitative indicator of fluctuation amplitude.

[0053] The load change trend is determined through linear regression slope analysis or first-order difference sign determination. As an example, this disclosure uses the least squares method to linearly fit the predicted sequence to obtain the slope. Based on the sign and magnitude of the slope and a preset trend discrimination threshold, the load change trend is classified as an upward trend, a downward trend, or a stable trend.

[0054] Based on load fluctuations and load change trends, the low-load trigger threshold and high-load trigger threshold are dynamically adjusted. As an example, this disclosure uses a preset mapping rule to achieve dynamic adjustment of the thresholds, specifically including the following scenarios: When the load fluctuation is less than the first threshold (e.g., 10%), it is determined to be a stable load scenario. The low load trigger threshold is set to the first low value (e.g., 25%), and the high load trigger threshold is set to the first high value (e.g., 80%). When the load fluctuation is greater than or equal to the second threshold (e.g., 15%), it is judged as a violent fluctuation scenario. The low load trigger threshold is set to the second low value (e.g., 35%), and the high load trigger threshold is set to the second high value (e.g., 70%). The second low value is higher than the first low value, and the second high value is lower than the first high value, in order to avoid frequent node state switching due to instantaneous spikes. When the load change trend is downward, the low load determination is triggered after a preset first time window (e.g., 3 minutes) to prevent missing the critical point due to a short-term load drop. When the load trend is upward, the high load trigger threshold is lowered by a preset percentage (e.g., 10%) in advance to make it easier to meet the expansion conditions and reserve response time for the upcoming high load.

[0055] After the above dynamic adjustments, the thresholds for the current decision cycle to take effect are obtained: low load trigger threshold and high load trigger threshold.

[0056] Obtain real-time load data for each node. Real-time load data originates from the load values ​​collected in real-time by the multimodal acquisition agents deployed on each node. It should be noted that real-time load data can be a single-dimensional comprehensive load metric (such as CPU utilization) or a comprehensive load value fused from multiple modalities.

[0057] Based on the effective low-load trigger threshold, high-load trigger threshold, and real-time load data, the load status of each node is determined. As an example, this disclosure adopts the following comprehensive judgment rule: Low load state: A low load state is defined as when the real-time load data is less than or equal to the low load trigger threshold and the load change trend indicates that there are no signs of load recovery within a preset second time window (e.g., 5 minutes). The absence of signs of load recovery can be determined by whether the values ​​of subsequent time steps in the prediction sequence remain below the low load trigger threshold and the slope is not positive.

[0058] High load state: A state is determined to be high load when any of the following conditions are met: Real-time load data is greater than or equal to the high load trigger threshold; The load change trend indicates that a high load state is about to be reached, and the maximum value in the predicted sequence is about to exceed the high load trigger threshold.

[0059] Normal state: If none of the above conditions are met, it is determined to be a normal state, and the node maintains the current operating mode.

[0060] Based on the above-described determination rules, this disclosure can output a node status list containing node identifiers and their corresponding states. This node status list can serve as input for the automatic execution of subsequent energy-saving strategies, triggering corresponding energy-saving or capacity expansion operations.

[0061] Cloud computing energy-saving solutions in related technologies typically use fixed thresholds (e.g., low load threshold = 30%, high load threshold = 75%) to trigger energy-saving operations. This rigid threshold setting cannot adapt to the dynamic fluctuation characteristics of load in different business scenarios such as web services, big data processing, and deep learning training. As a result, energy-saving or expansion operations may be accidentally triggered in scenarios with drastic fluctuations, while energy-saving opportunities may be missed in stable scenarios.

[0062] To address the aforementioned issues, this disclosure employs a load trend-driven dynamic threshold adjustment algorithm. Based on the load prediction value output by LSTM, the algorithm analyzes the load fluctuation amplitude and trend, dynamically adjusts the low load trigger threshold and high load trigger threshold according to preset mapping rules, and integrates the adjusted thresholds with real-time load data to comprehensively determine the node load status.

[0063] Therefore, this disclosure achieves adaptive adjustment of the threshold, which can automatically adapt to different business scenarios according to the fluctuation characteristics of the load, maximize energy saving opportunities while ensuring service quality, and effectively solve the misjudgment problem caused by fixed thresholds.

[0064] Step 140: For any of the nodes, in response to the node's load state being low load, perform an energy-saving operation on the node in the low load state; in response to the node's load state being high load, perform a capacity expansion operation on the node in the high load state.

[0065] In some exemplary embodiments, the step of performing energy-saving operations on nodes in the low-load state in response to the node's load state being a low-load state includes: Cache the image of the operating system and application environment of the node in the low load state locally; A soft shutdown command is sent to the node in the low load state through the intelligent platform management interface, causing the node in the low load state to enter a low-power standby state that retains power only from the baseboard management controller.

[0066] When a node's load status is marked as low load, the idle node filtering module first further filters the nodes marked as low load to generate an idle node list. The idle node filtering module uses the following composite judgment rules: CPU utilization <5%, memory utilization <10%, and the load prediction value indicates no upward trend in load within the next 5 minutes.

[0067] Image Cache: Triggering the image cache module quickly compresses the operating system state, application environment, and memory data of the idle node into an image file and stores it on the node's local SSD (Solid State Drive). This image file is used for direct loading during subsequent fast boots, skipping the complete operating system startup process.

[0068] Soft shutdown execution: A soft shutdown command (e.g., ACPI soft shutdown) is sent to idle nodes via IPMI (Intelligent Platform Management Interface) to orderly shut down the operating system and cut off the main power supply, leaving only the BMC (Baseboard Management Controller) powered. At this time, the node enters a low-power standby state with power consumption below 5W, significantly reducing energy consumption. Status logging: The soft shutdown execution result (success / failure) and the timestamp of the node entering sleep mode are recorded as feedback information for subsequent use.

[0069] In some exemplary embodiments, the scaling operation performed on the node in the high-load state in response to the node's load state includes: The hibernation nodes are sorted based on hardware performance indicators, and the hibernation node with the best hardware performance is selected as the wake-up target node. The system sends a power-on command to the wake-up target node through the intelligent platform management interface and configures the pre-boot execution environment to enable the wake-up target node to load the locally cached image, skip the hardware self-test process, complete the power-on, and automatically connect to the resource pool. Migrate the application or virtual machine on the node that is in a high-load state to the wake-up target node.

[0070] Specifically, when a node's load status is marked as high load, the Fast Boot Manager immediately initiates the expansion process. To ensure the performance and response speed of the wake-up nodes, this disclosure employs the following steps: Node sorting and selection: Sort the list of currently sleeping nodes in descending order according to hardware performance indicators (such as CPU frequency, number of cores, and memory capacity) to generate a wake-up priority queue. The node with the highest score is selected as the wake-up target first.

[0071] Fast boot execution: A power-on command is sent to the selected hibernation node via IPMI, and PXE (Preboot Execution Environment) boot parameters are configured to enable the node to boot from the network. The configuration skips the full memory hardware self-test (POST) process, directly loading the image file previously cached in the local SSD via PXE. This image loading process directly restores the operating system and application environment to memory, reducing node wake-up time from several minutes in traditional methods to tens of seconds.

[0072] Automatic access to the resource pool: After the node completes the fast boot, it automatically sends registration information to the cluster management system, joins the computing resource pool, and reports its available resource status (such as remaining CPU, memory, etc.).

[0073] Load balancing scheduling: The load balancer receives resource status information from the woken-up nodes and generates VM (virtual machine) migration instructions according to a preset scheduling policy. The scheduling policy can employ a resource utilization balancing algorithm to migrate some virtual machines or containers on high-load nodes to newly woken-up nodes. The selection of the migration target node is based on the principle of minimizing the variance of cluster resource utilization.

[0074] In related cloud computing energy-saving solutions, when a high load is detected requiring the waking of a dormant node, a power-on command is typically sent via IPMI. This method requires the node to undergo a complete hardware self-test, operating system startup, and application loading process, which can take several minutes to wake up. This approach cannot respond promptly to sudden high loads and may lead to service level agreement violations, affecting service availability.

[0075] To address the aforementioned issues, this disclosure employs a collaborative mechanism of soft shutdown and fast startup. Before performing a soft shutdown on a low-load node, the node's operating system and application environment image is cached to the local SSD. When the node needs to be woken up, a power-on command is sent via IPMI and combined with PXE boot technology to directly load the locally cached image, skipping the hardware self-test process to quickly complete the startup and access the resource pool.

[0076] Therefore, this disclosure reduces the node wake-up time from several minutes to tens of seconds, significantly improving the response speed to sudden loads and meeting the stringent service quality requirements in high availability scenarios. At the same time, by using a load balancer to migrate applications on high-load nodes to wake-up nodes, dynamic optimization of cluster resources is achieved.

[0077] In some exemplary embodiments, the method further includes: The execution results of step 140 are fed back to steps 110 and 120 respectively, which are used to adjust the data acquisition frequency and optimize the parameters of the LSTM prediction model to form a closed-loop control.

[0078] Specifically, the data acquisition density in step 110 is adjusted based on the frequency of high-load events in the execution results; the execution results are used as new training samples to incrementally train the LSTM prediction model, optimizing attention weights and prediction accuracy.

[0079] In step 110, the frequency of multimodal data collection can be adaptively adjusted based on the severity of load fluctuations reflected in the execution results. When the execution results indicate frequent high-load events (e.g., the number of rapid system startups exceeding a threshold within a unit of time) or significant fluctuations in node utilization after load balancing, step 110 determines that the current scenario is a high-dynamic load scenario and shortens the data collection interval from the default value (e.g., 1 second) to (e.g., 0.5 seconds) to capture more granular load changes. Conversely, if no energy-saving or expansion operations are triggered for a long time and the node status is stable, the collection interval is extended (e.g., 2 seconds) to reduce system overhead.

[0080] Step 120 utilizes the execution results to incrementally train and optimize the parameters of the LSTM prediction model, enabling the model to continuously adapt to the current load pattern. Specifically, step 120 compares the actual load data collected after step 140 with the previously output predicted values ​​from step 120 to calculate the prediction error. These error samples are stored in an experience replay buffer. When the buffer accumulates a sufficient number of new samples, incremental training of the model is triggered.

[0081] Incremental training uses mini-batch gradient descent to update the parameters of the LSTM network and attention mechanism. The gradient is calculated and the weight parameters are updated using the backpropagation algorithm.

[0082] Furthermore, the energy-saving effect reflected in the execution results (such as actual energy savings) can also serve as a reinforcement learning signal to adjust the initial weight allocation of different modalities in the attention mechanism, making the model pay more attention to the modalities that contribute more to the prediction of the current scene. For example, if multiple consecutive prediction errors are mainly due to misjudgments of network load, the base weight of the network modalities in the attention mechanism can be appropriately increased.

[0083] Through the aforementioned feedback optimization mechanism, the LSTM prediction model in step 120 can track changes in load patterns in real time, avoiding the decline in prediction accuracy caused by model aging, thereby continuously providing high-quality load prediction values ​​for step 130. Simultaneously, the adaptive adjustment of the acquisition frequency in step 110 ensures a balance between the timeliness of input data and resource consumption, ultimately forming an intelligent closed-loop management system of acquisition, prediction, decision-making, execution, and optimization. This enables the system to dynamically adapt to load changes under different business scenarios, achieving a long-term optimal balance between energy efficiency and service quality.

[0084] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0085] The cloud computing facility control method provided in the above embodiments can be implemented as a cloud computing facility control system on the product side.

[0086] Figure 2 This is a schematic diagram of the structure of a cloud computing facility control system provided in an embodiment of this disclosure.

[0087] The cloud computing facility control system includes a perception layer, an LSTM prediction layer, a dynamic threshold decision layer, and an execution layer.

[0088] In the perception layer, it is used for multi-dimensional data acquisition and preprocessing.

[0089] The perception layer deploys multimodal acquisition nodes, each configured with various heterogeneous sensors, including CPU utilization sensors, memory usage sensors, network bandwidth utilization sensors, disk I / O activity sensors, and device surface temperature sensors. These sensors collect raw hardware and resource status data from each node in real time. The collected raw data is input to a normalized feature vector module for processing. This module first uses the 3σ principle to remove outliers, then eliminates dimensional differences through Min-Max normalization, and finally outputs a standardized "60-step historical feature vector." This historical feature vector has a dimension of 60x5, corresponding to five-dimensional multimodal data with one sampling point per second over the past 60 seconds, and serves as the input to the LSTM prediction layer.

[0090] In the LSTM prediction layer, it is used for load trend prediction and feature weight allocation.

[0091] The prediction layer contains an LSTM ensemble model and an attention mechanism module.

[0092] The LSTM ensemble model consists of three stacked LSTM networks: 64 neurons, 128 neurons, and 64 neurons respectively. This model receives a 60-step historical feature vector from the perception layer, extracts temporal features through a gating mechanism, and generates a hidden state sequence. The attention mechanism module receives this hidden state sequence and calculates the relevance scores of each time step and modality feature to the current prediction task using additive attention. After softmax normalization, the attention weights for each modality feature are obtained, achieving dynamic allocation of feature weights. The attention mechanism module dynamically allocates weights for each dimension based on the importance of the current load features (e.g., assigning higher weights to the CPU modality in CPU-intensive tasks). Finally, the weighted comprehensive feature representation is input into the fully connected layer, outputting a "10-step load prediction value," corresponding to the load change trend over the next 10 time steps (30 seconds per step, totaling 5 minutes). Simultaneously, the prediction layer generates initial suggested values ​​for low-load and high-load trigger thresholds based on the current load features, which are then passed to the dynamic threshold decision layer.

[0093] In the dynamic threshold decision layer, it is used for load status determination and node selection.

[0094] The dynamic threshold decision layer includes a threshold adjustment module, a low load determination module, and an idle node list module.

[0095] The threshold adjustment module receives 10-step load prediction values ​​from the prediction layer. First, it determines the load fluctuation amplitude by calculating the standard deviation of the prediction sequence, and then determines the load change trend (increasing / decreasing / stable) through linear regression slope analysis. Based on the load fluctuation amplitude and load change trend, the threshold adjustment module dynamically adjusts the specific values ​​of the low-load trigger threshold and the high-load trigger threshold according to preset mapping rules: for example, when the load is stable, the low-load trigger threshold is set to 25% and the high-load trigger threshold to 80%; when the load fluctuates greatly, the low-load trigger threshold is set to 35% and the high-load trigger threshold to 70%; when the load is decreasing, a soft shutdown is triggered 3 minutes later; when the load is increasing, the high-load trigger threshold is lowered by 10% 5 minutes earlier. The adjusted effective thresholds are then passed to the low-load determination module.

[0096] The low-load determination module integrates real-time load data with 10-step load prediction values, comparing them against effective low-load and high-load trigger thresholds to determine the node status. Specifically, a node is marked as low-load if it consistently meets the condition of "predicted load ≤ low-load trigger threshold" for 5 minutes; and as high-load if "predicted load ≥ high-load trigger threshold". The idle node list module further filters the nodes marked as low-load, using a composite rule of "CPU < 5% + memory < 10% + predicted no load increase in the next 5 minutes" to generate an "idle node list". This list, along with the high-load trigger signal, is then transmitted to different components in the execution layer.

[0097] In the execution layer, it is used to automate soft shutdown, fast startup, and load balancing.

[0098] The execution layer includes a soft shutdown manager, a fast startup manager, and a load balancer scheduler.

[0099] The soft shutdown manager receives a list of idle nodes, first triggering the image caching module to cache the operating system and application environment of the idle nodes to the local SSD; then it sends a soft shutdown signal to the idle nodes via IPMI, causing the nodes to enter a low-power standby state (power consumption <5W) with only BMC power supply.

[0100] Upon receiving a high-load trigger signal, the Fast Boot Manager immediately initiates the expansion process. First, it sorts the hibernating nodes by hardware performance (such as CPU frequency and memory capacity), selecting the node with the best performance as the wake-up target. Then, it sends a power-on command via IPMI, and in conjunction with the PXE boot controller, enables the wake-up target to directly load the cached image from its local SSD, skipping the hardware self-test process to quickly complete booting and automatically connect to the resource pool. Finally, it feeds back the resource status of the wake-up node to the load balancer.

[0101] The load balancer scheduler receives the resource status of the wake-up node, generates VM migration instructions, and migrates applications or virtual machines on high-load nodes to the wake-up node, ensuring that the cluster resource utilization remains stable within a stable range and completing load balancing.

[0102] To implement the above embodiments, this disclosure also proposes a cloud computing facility control device.

[0103] Figure 3 This is a schematic diagram of a cloud computing facility control device provided in an embodiment of this disclosure. The device can be implemented by software and / or hardware, and is generally integrated into an electronic device. Figure 2 As shown, the device includes: a multimodal operation data determination module 210, a load prediction value determination module 220, a load status determination module 230, and a control execution module 240, wherein: The multimodal operation data determination module 210 is configured to acquire multimodal operation data of each node of the cloud computing facility based on multiple heterogeneous types of sensors, preprocess the multimodal operation data, and obtain a standardized historical feature vector. The load prediction value determination module 220 is configured to input the standardized historical feature vector into a pre-trained long short-term memory network ensemble model, and dynamically allocate the weights of the features of each modality through the pre-trained long short-term memory network ensemble model combined with an attention mechanism to predict the load prediction values ​​for multiple future time steps. The load status determination module 230 is configured to determine the load fluctuation amplitude and load change trend based on the load prediction value, dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation amplitude and the load change trend, obtain real-time load data of each node, and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold and the real-time load data. The control execution module 240 is configured to perform energy-saving operations on nodes in a low-load state in response to the load state of the node being low-load, and to perform capacity expansion operations on nodes in a high-load state in response to the load state of the node being high-load.

[0104] In some exemplary embodiments, the load prediction value determination module 220 is specifically configured as follows: The standardized historical feature vector is input into the pre-trained long short-term memory network ensemble model, and the temporal features are extracted through the pre-trained long short-term memory network ensemble model to generate a hidden state sequence. The hidden state sequence is input into the attention mechanism. Based on the attention mechanism, the correlation scores between each time step and each modality feature and the current prediction task are calculated. After normalization, the attention weights of each modality feature are obtained. The attention weights are summed with their corresponding hidden states to obtain the context vector. The context vector is input into the fully connected layer, which outputs the load prediction values ​​for multiple future time steps.

[0105] In some exemplary embodiments, the load state determination module 230 is specifically configured as follows: A load forecast value sequence is constructed based on the load forecast value, and statistical feature calculations are performed based on the load forecast value sequence to determine the load fluctuation amplitude and the load change trend. Using preset mapping rules, the low load trigger threshold and the high load trigger threshold are dynamically adjusted based on the load fluctuation amplitude and the load change trend; The real-time load data is compared with the low load trigger threshold and the high load trigger threshold to determine the load state of each node, which includes a low load state and a high load state.

[0106] In some exemplary embodiments, the control execution module 240 is specifically configured as follows: Cache the image of the operating system and application environment of the node in the low load state locally; A soft shutdown command is sent to the node in the low load state through the intelligent platform management interface, causing the node in the low load state to enter a low-power standby state that retains power only from the baseboard management controller.

[0107] In some exemplary embodiments, the control execution module 240 is specifically configured as follows: The hibernation nodes are sorted based on hardware performance indicators, and the hibernation node with the best hardware performance is selected as the wake-up target node. The system sends a power-on command to the wake-up target node through the intelligent platform management interface and configures the pre-boot execution environment to enable the wake-up target node to load the locally cached image, skip the hardware self-test process, complete the power-on, and automatically connect to the resource pool. Migrate the application or virtual machine on the node that is in a high-load state to the wake-up target node.

[0108] In some exemplary embodiments, the multimodal operational data includes: CPU utilization, memory usage, network bandwidth utilization, disk I / O activity, and device surface temperature.

[0109] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0110] The cloud computing facility control device provided in this disclosure can execute the cloud computing facility control method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of executing the method.

[0111] To implement the above embodiments, this disclosure also proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implements the cloud computing facility control method in the above embodiments.

[0112] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure.

[0113] The following is a detailed reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device 300 in the embodiments of this disclosure. The electronic device 300 in the embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0114] like Figure 4 As shown, the electronic device 300 may include a processor (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a memory 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0115] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0116] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from memory 308, or installed from ROM 302. When the computer program is executed by processor 301, it performs the functions defined in the cloud computing facility control method of embodiments of this disclosure.

[0117] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0118] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0119] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0120] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the aforementioned cloud computing facility control method.

[0121] Electronic devices can be programmed with computer program code in one or more programming languages ​​or a combination thereof to perform the operations of this disclosure. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0122] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0123] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0124] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0125] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0126] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0127] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0128] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A cloud computing facility control method characterized by comprising: The method includes: Multimodal operation data of each node of the cloud computing facility is acquired based on various heterogeneous types of sensors, and the multimodal operation data is preprocessed to obtain a standardized historical feature vector. The standardized historical feature vectors are input into a pre-trained long short-term memory network ensemble model. The pre-trained long short-term memory network ensemble model, combined with an attention mechanism, dynamically allocates the weights of features of each modality and predicts the load prediction values ​​for multiple future time steps. Based on the load forecast value, determine the load fluctuation range and load change trend, dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation range and load change trend, obtain real-time load data of each node, and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold and the real-time load data. For any of the nodes, in response to the node's load state being low load, an energy-saving operation is performed on the node in the low load state; in response to the node's load state being high load, a capacity expansion operation is performed on the node in the high load state.

2. The method of claim 1, wherein, The process involves inputting the standardized historical feature vector into a pre-trained Long Short-Term Memory (LSTM) network ensemble model. Through this model, combined with an attention mechanism, the weights of features from each modality are dynamically allocated to predict the load forecast values ​​for multiple future time steps. This includes: The standardized historical feature vector is input into the pre-trained long short-term memory network ensemble model, and the temporal features are extracted through the pre-trained long short-term memory network ensemble model to generate a hidden state sequence. The hidden state sequence is input into the attention mechanism. Based on the attention mechanism, the correlation scores between each time step and each modality feature and the current prediction task are calculated. After normalization, the attention weights of each modality feature are obtained. The attention weights are summed with the corresponding hidden states to obtain the context vector; The context vector is input into the fully connected layer, which outputs the load prediction values ​​for multiple future time steps.

3. The method of claim 1, wherein, The process of determining the load fluctuation amplitude and load change trend based on the load forecast value, dynamically adjusting the low load trigger threshold and high load trigger threshold based on the load fluctuation amplitude and load change trend, acquiring real-time load data for each node, and determining the load status of each node based on the low load trigger threshold, the high load trigger threshold, and the real-time load data, includes: A load forecast value sequence is constructed based on the load forecast value, and statistical feature calculations are performed based on the load forecast value sequence to determine the load fluctuation amplitude and the load change trend. Using preset mapping rules, the low load trigger threshold and the high load trigger threshold are dynamically adjusted based on the load fluctuation amplitude and the load change trend; The real-time load data is compared with the low load trigger threshold and the high load trigger threshold to determine the load state of each node, which includes a low load state and a high load state.

4. The method of claim 1, wherein, The response to the node's load state being a low-load state, performing energy-saving operations on the node in the low-load state, includes: Cache the image of the operating system and application environment of the node in the low load state locally; A soft shutdown command is sent to the node in the low load state through the intelligent platform management interface, causing the node in the low load state to enter a low-power standby state that retains power only from the baseboard management controller.

5. The method of claim 1, wherein, The response to the node's load state being high load, performing a scaling operation on the node in the high load state, includes: The hibernation nodes are sorted based on hardware performance indicators, and the hibernation node with the best hardware performance is selected as the wake-up target node. The system sends a power-on command to the wake-up target node through the intelligent platform management interface and configures the pre-boot execution environment to enable the wake-up target node to load the locally cached image, skip the hardware self-test process, complete the power-on, and automatically connect to the resource pool. Migrate the application or virtual machine on the node that is in a high-load state to the wake-up target node.

6. The method of claim 1, wherein, The multimodal operation data includes: CPU utilization, memory usage, network bandwidth utilization, disk I / O activity, and device surface temperature.

7. A cloud computing facility control device characterized by comprising: The device includes: The multimodal operation data determination module is configured to acquire multimodal operation data of each node of the cloud computing facility based on multiple heterogeneous types of sensors, preprocess the multimodal operation data, and obtain a standardized historical feature vector. The load prediction value determination module is configured to input the standardized historical feature vector into a pre-trained long short-term memory network ensemble model, and dynamically allocate the weights of the features of each modality through the pre-trained long short-term memory network ensemble model combined with an attention mechanism to predict the load prediction values ​​for multiple future time steps. The load status determination module is configured to determine the load fluctuation amplitude and load change trend based on the load prediction value, dynamically adjust the low load trigger threshold and high load trigger threshold based on the load fluctuation amplitude and the load change trend, obtain real-time load data of each node, and determine the load status of each node based on the low load trigger threshold, the high load trigger threshold and the real-time load data. The control execution module is configured to, for any of the nodes, perform energy-saving operations on nodes in a low-load state in response to the node's load state being low-load, and perform capacity expansion operations on nodes in a high-load state in response to the node's load state being high-load.

8. An electronic device, comprising: The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program / instruction thereon, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 6.

Images