Data center-oriented control method

Abstract

The invention discloses a data center-oriented control method. The method comprises the following steps: acquiring environment data of a data center acquired by multiple sensors in multiple directions of the data center; and inputting the environment data into the constructed fan rotating speed control model based on the neural network, and outputting the target rotating speed of each fan of the data center by the fan rotating speed control model for controlling the rotating speed of each fan of the data center. The method further comprises the following steps: recording the sizes of current idle resources of all servers of the data center, and sorting the idle resource blocks from small to large; receiving a resource request sent after an external service request enters the data center; and the idle resource table is searched in sequence, and the searched first idle resource block with the size greater than or equal to the request quantity is allocated to the external service requester. And updating the idle resource table, and feeding back the temperature of the server case and the rotating speed of the fan under the current resource occupation condition to the fan rotating speed control model. According to the invention, the operation stability of the data center server can be improved.

Description

Technical Field

[0001] The invention belongs to the field of automatic control, and in particular relates to a control method for a data center. Background Art

[0002] Data centers serve as "transit stations" for Internet information. They are responsible for transmitting, accelerating, and storing information for Internet network infrastructure. Data centers not only provide basic services such as information storage, recovery and backup information management, and networks, but also connect two or more data centers through high-speed packet communication technology. The huge amount of data processing makes the stable operation of data centers particularly important, so there are strict requirements for the environmental monitoring and heat dissipation capabilities of data centers. In the development of heat dissipation technology, the continuous high-load service of data center servers has brought high heat dissipation costs. According to the "2023 China Liquid Cooling Application Market Research Report", the heat dissipation cost of 1kW for a liquid-cooled data center in 2023 will be nearly 6,500 yuan.

[0003] As the application of integrated air-ground information networks in the civil field becomes more and more widespread, the increasing amount of data has put forward higher requirements on the data processing capabilities of ground communication network data centers. However, the current air-cooling technology has a single operation method during the application process, and only schedules the air-cooling system of data centers and servers according to changes in environmental conditions. There are problems such as low heat dissipation efficiency and ineffective heat dissipation of servers and other equipment. At the same time, the current data center has the problem of unreasonable server resource scheduling, which causes some servers to work at high load while some servers are idle, further affecting the effective heat dissipation of the data center. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a control method for a data center, which can improve the operation stability of a server in the data center.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] In a first aspect, the present application provides a control method for a data center, including controlling the rotation speed of each fan in the data center by the following method:

[0007] Obtain environmental data of the data center collected by multiple sensors at multiple locations in the data center;

[0008] The acquired environmental data is input into a constructed neural network-based fan speed control model, and the fan speed control model outputs a target speed of each fan in the data center, which is used to control the speed of each fan in the data center; wherein the constructed neural network-based fan speed control model has the function of outputting the target speed of each fan in the data center according to the input environmental data.

[0009] In a possible implementation, the multiple sensors include a temperature sensor, a humidity sensor, and a wind speed sensor.

[0010] In a possible implementation, before inputting the collected environmental data into the constructed neural network-based fan speed control model, the method further includes:

[0011] Before inputting the acquired environmental data into the constructed neural network-based fan speed control model, it also includes:

[0012] The collected environmental data are preprocessed, including using activation functions to extract the nonlinear features of the environmental data and normalizing each environmental data time series.

[0013] In one possible implementation, the fan speed control model includes an intramodal attention layer, a cross-modal attention layer, and a fully connected layer:

[0014] The intramodal attention layer is used to learn the time series information within each modality. The calculation method includes:

[0015] For the i-th environmental data X i ∈R 1*d , calculate its internal modal attention score matrix A i =softmax(W ai X i +b ai ), where d is the environmental data X i The number of acquisition time points, W ai ∈R d*1 is the weight matrix of linear transformation, b ai is the bias term, both of which are optimized during the training process. i ∈R d*d According to the calculation result of the intramodal attention score matrix, we can get the intramodal weighted feature representation of the i-th environment data: Y i =A i W ci , where W ci ∈R d*1 is the transformation matrix, Y i ∈R d*1 The Y i W is the feature output of the i-th environment data through the intramodal attention layer. ai , b ai , W ci All are optimized during the training process.

[0016] The cross-modal attention layer is used to learn the relationship between different modalities. The calculation method includes:

[0017] For two different environment data X i ∈R 1*d , X j ∈R 1*d , calculate the score Among them, A ij ∈R d*d is the attention weight matrix of the j-th environment data to the i-th environment data, is the scaling factor, which can be an empirical value.

[0018] According to the calculation results of the attention weight matrix, the cross-modal weighted feature representation of the i-th environment data can be obtained: in, n is the number of modes, that is, the number of types of environmental data. is the feature output of the i-th environment data through the cross-modal attention layer.

[0019] The fully connected layer is used to output the fan target speed prediction result based on the learning results of the intramodal attention layer and the cross-modal attention layer. The calculation method includes:

[0020] Set Y1, Y2, ..., Y n Combine to form Y,

[0021] Will Merge, constitute

[0022] According to X, Y, Get feature representation for decision making Where X is the environmental data matrix, and its i-th row is the environmental data X i , X∈R n*d .

[0023] Calculate the decision variable M = (ZV) T t, where V∈R n*m , t∈R d*1 , M∈R m*1 , m is the number of fans in the data center, and M is used to represent the rotation speed of each fan in the data center.

[0024] In a possible implementation, the method further includes performing task scheduling on the data center using the following method:

[0025] Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center (including available CPU, memory, and disk resources);

[0026] Request processing: receiving resource requests issued by external service requests after entering the data center. The resource request specifies the size of resources (CPU, memory, and disk resources) required by the external service request, i.e., the request amount;

[0027] Resource search and allocation: traverse the idle resource table and check whether the current idle resources of each server can meet the needs of the external service request;

[0028] If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated;

[0029] Idle resource table update: recalculate the resource usage of the server and update the idle resource table to reflect the latest idle resource situation.

[0030] In a possible implementation, after the idle resource table is updated, the temperature and fan speed of the current server chassis under the current resource occupancy condition are fed back to the fan speed control model in real time.

[0031] In a possible implementation, the idle resources of the server are divided into a plurality of idle resource blocks, and the sizes of the plurality of idle resource blocks are different;

[0032] In the idle resource table, all idle resource blocks are sorted from small to large according to their sizes;

[0033] This step can form an ordered list of idle resource amounts from small to large, which is convenient for finding and allocating the smallest idle resource block that meets the request, thereby avoiding wasting larger idle resource blocks.

[0034] In the resource search and allocation, a first idle resource block is searched, wherein the first idle resource block is a first idle resource block whose size is greater than or equal to the requested amount and is found by traversing the idle resource table from front to back;

[0035] If the size of the first idle resource block is equal to the requested amount, the entire first idle resource block is fully allocated to the external service requester;

[0036] If the size of the first idle resource block is larger than the requested amount, the portion of the first idle resource block equal to the requested amount will be allocated to the external service requester, and the remaining portion will still be retained as an idle resource block in the idle resource table, and its size will be reduced accordingly, that is, equal to the original size minus the requested amount.

[0037] In a second aspect, the present application provides a control system for a data center, including an intelligent data analysis module;

[0038] The intelligent data analysis module is used to obtain environmental data of the data center collected by various sensors in multiple locations of the data center; the obtained environmental data is input into a constructed neural network-based fan speed control model, and the fan speed control model outputs a target speed of each fan in the data center, which is used to control the speed of each fan in the data center; wherein the constructed neural network-based fan speed control model has the function of outputting the target speed of each fan in the data center according to the input environmental data.

[0039] In a possible implementation, the system further includes a task scheduling module for task scheduling, including:

[0040] Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center (including available CPU, memory, and disk resources);

[0041] Request processing: receiving resource requests issued by external service requests after entering the data center. The resource request specifies the size of resources (CPU, memory, and disk resources) required by the external service request, i.e., the request amount;

[0042] Resource search and allocation: traverse the idle resource table and check whether the current idle resources of each server can meet the needs of the external service request;

[0043] If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated;

[0044] Idle resource table update: recalculate the resource usage of the server and update the idle resource table to reflect the latest idle resource situation.

[0045] In a third aspect, the present application provides an electronic device, including: a memory and a processor;

[0046] The memory is used to store computer programs;

[0047] The processor is used to call the computer program to execute the method as described above.

[0048] In a fourth aspect, the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed on an electronic device, the electronic device implements the method described above.

[0049] In a fifth aspect, the present application provides a computer program product, including a computer program, and when the computer program is run on an electronic device, the electronic device implements the method as described above.

[0050] The specific implementation methods of the second to fifth aspects of the present application can refer to the implementation methods of the first aspect, which will not be repeated here.

[0051] Beneficial effects:

[0052] This application adjusts the fan speed of the data center according to multimodal time series, including changes in the data center ambient temperature, humidity, air flow rate, and temperature of the data center servers, in order to provide the data center and its internal servers with the most efficient cooling solution and realize intelligent environmental management. At the same time, the task scheduling strategy designed in this application uses the optimal adaptation algorithm to sort and allocate server idle resources based on the hardware information of each server, the CPU occupancy rate, memory occupancy rate, and hard disk usage of the comprehensive server, so as to achieve effective regulation and control of server resources and dynamic allocation of tasks, so as to avoid uneven resource allocation, resulting in some servers being heavily loaded while some servers have more idle resources. Combining the above methods, the stability of the data center server can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a system framework flow chart in the embodiment of this application;

[0054] Figure 2 This is a framework diagram of the task scheduling system in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The present application is further described below in conjunction with the accompanying drawings and specific embodiments.

[0056] like Figure 1 and Figure 2As shown, the present application discloses a control method and system for data centers, which can perform intelligent environmental management and dynamic task allocation based on multimodal time series. The corresponding system may include an intelligent data analysis module and a task intelligent dynamic scheduling strategy module; multimodal signals (including humidity, temperature, wind speed and other signals) collected by various sensors are input into the intelligent data analysis module; the intelligent data analysis module can perform standardized processing (including normalization processing, etc.) on the collected humidity, temperature, wind speed and other signals, and innovatively integrate multimodal information to comprehensively capture various factors affecting the fan speed. A heterogeneous network graph can be constructed using multimodal information input, and an intelligent algorithm, including an internal modal attention layer and a cross-modal source attention layer, is used to learn the relationship between different modal information and fan speed control under different time states, so as to efficiently and accurately predict and realize real-time control of fan speed. At the same time, by obtaining data information such as CPU occupancy, memory utilization, and disk remaining space, a task scheduling strategy is adopted to realize dynamic allocation of tasks to ensure that the data center is in the best operating state. In addition, the operating status of the data center is fed back to the intelligent data analysis module in real time to realize feedback control. At the same time, the data center operation status can also be displayed in real time through the data center interactive window (such as the data center browser) so that the staff can master and control it. This application greatly improves the intelligent control level of the product in actual use status and provides more information reference for task allocation.

[0057] The present application is described in further detail below.

[0058] A control method for a data center obtains environmental data of the data center collected by multiple sensors at multiple locations in the data center; the obtained environmental data is input into a constructed fan speed control model based on a neural network, and the fan speed control model outputs the target speed of each fan in the data center to control the speed of each fan in the data center to ensure maximum heat dissipation efficiency.

[0059] The various sensors include a temperature sensor, a humidity sensor, and a wind speed sensor.

[0060] Sensors in multiple locations, such as temperature sensors, may include an outdoor temperature sensor and a chassis temperature sensor for respectively collecting the external ambient temperature of the data center and the chassis temperature of the server in the data center.

[0061] The fans in the data center include fans in the air cooling system of the data center and fans in the server chassis.

[0062] The data center collects various environmental data through temperature sensors, humidity sensors, wind speed sensors, etc. in multiple locations of the data center. The collected environmental data can be transmitted to the intelligent data analysis module through the wireless network;

[0063] In some embodiments, before inputting the acquired environmental data into the constructed neural network-based fan speed control model, the method further includes:

[0064] The intelligent data analysis module pre-processes the collected environmental data.

[0065] In some embodiments, the preprocessing includes extracting nonlinear features of the environmental data using, for example, a ReLU activation function.

[0066] In some embodiments, the preprocessing includes normalizing each environmental data time series.

[0067] The above-mentioned various environmental data include temperature, humidity, wind speed, etc. at multiple locations of the current data center.

[0068] A variety of environmental data is a source of information in multiple modes. Each mode corresponds to a type of environmental data. The variables collected by each sensor affect each other and jointly affect the air cooling system. The intelligent data analysis module can learn the relationship between the fan speed and various environmental data through the fan speed control model based on the neural network, so that various environmental data converge to the most ideal value under appropriate fan speed control, so as to combine various environmental data of the current data center, predict the fan target speed, and control the fan speed of the data center according to the fan target speed.

[0069] In some embodiments, the fan speed control model includes an intramodal attention layer, a cross-modal attention layer, and a fully connected layer;

[0070] The intramodal attention layer is used to learn the time series information within each modality;

[0071] The purpose of introducing the intramodal attention layer is to learn the temporal relationship of each environmental data, use the self-attention mechanism to interact information within each modality, and generate weighted representations to learn the impact of the same environmental data at different times.

[0072] For the i-th environmental data X i ∈R 1*d , calculate its internal modal attention score matrix A i =softmax(W ai X i +b ai ), where d is the environmental data X i The number of acquisition time points, W ai ∈R d*1 is the weight matrix of linear transformation, b ai is the bias term, both of which are optimized during the training process. i ∈R d*dAccording to the calculation result of the intramodal attention score matrix, we can get the intramodal weighted feature representation of the i-th environment data: Y i =A i W ci , where W ci ∈R d*1 is the transformation matrix, Y i ∈R d*1 The Y i W is the feature output of the i-th environment data through the intramodal attention layer. ai , b ai , W ci All are optimized during the training process.

[0073] The cross-modal attention layer is used to learn the relationship between different modalities;

[0074] The purpose of introducing the cross-modal attention layer is to learn the relationship between different environmental data and strengthen the interdependence between temperature, humidity, and air flow rate.

[0075] For two different environment data X i ∈R 1*d , X j ∈R 1*d , calculate the score Among them, A ij ∈R d*d is the attention weight matrix of the j-th environment data to the i-th environment data, is the scaling factor, which can be an empirical value.

[0076] According to the calculation results of the attention weight matrix, the cross-modal weighted feature representation of the i-th environment data can be obtained: in, n is the number of modes, that is, the number of types of environmental data. is the feature output of the i-th environment data through the cross-modal attention layer.

[0077] The fully connected layer is used to output the fan target speed prediction result based on the learning results of the intramodal attention layer and the cross-modal attention layer.

[0078] Set Y1, Y2, ..., Y n Merge to form Y, Y∈R d*n ;

[0079] Will Merge, constitute

[0080] According to X, Y, Get feature representation for decision making Where X is the environmental data matrix, and its i-th row is the environmental data X i , X∈R n*d .

[0081] According to the feature dimensions to be generated, calculate the decision variable M = (ZV) T t, where V∈R n*m , t∈R d*1 , M∈R m *1 , m is the number of fans in the data center, and M is used to represent the rotation speed of each fan in the data center.

[0082] This application uses the prediction of fan speed based on multiple environmental data as a multimodal processing task. First, the features of each environmental data are processed through the internal modal attention layer to obtain the internal modal weighted feature representation of each environmental data. Then, different environmental data are interacted and fused through the cross-modal attention layer to obtain the cross-modal weighted feature representation of each environmental data. Finally, the internal modal weighted feature representation and cross-modal weighted feature representation of each environmental data are fused to obtain the feature representation for decision-making, and then the decision variable, that is, the target speed of each fan in the data center, is calculated based on the feature representation for decision-making.

[0083] This application establishes data association relationships through the above neural network, taking into account the correlation between time points, so that the fusion of multiple environmental data no longer depends solely on the value of the current time point, but combines the influence of previous and next time points. Each row of the final result represents a new fusion feature, which weights different environmental data respectively and takes into account the changes between time steps.

[0084] In some embodiments, the preprocessed data is used to construct a heterogeneous network graph, in which the various environmental data are source nodes and the fan target speed is used as the predicted target node. The neural network is a graph neural network, which uses the heterogeneous network as input and predicts the value of the target node.

[0085] In some embodiments, the neural network is a back propagation neural network (BP neural network).

[0086] Design the reward function R(C) corresponding to the environmental data t , H t ,s t ), including deviation penalty and energy consumption penalty. Then the expression of the reward function is: Among them, E t is the energy consumption cost at time t, C i With X i，tare the environmental data under ideal conditions and current time conditions respectively, β i is the reward weight factor for the i-th environment data.

[0087] At time t, the fan target speed s is obtained through the fan speed control model. t Then, the fan speed of the data center is regulated according to the fan target speed. At time t+1, the new environmental data X is measured. i，t+1 , and store it. From this, the fan speed s executed at time t can be determined t The environmental changes brought about by the reward function R(E t , X i，t ) calculation results can be used to determine whether the current data center fan speed control strategy is conducive to making the environment closer to the ideal situation.

[0088] During the training of the neural network model, 70% of the data set is divided as a training set, 15% as a validation set, and 15% as a test set. The data set comes from the previous monitoring information of the data center. The training set is used to train the fan speed control model, the validation set is used to calculate the loss function value during the training of the fan speed control model, and the test set is the loss function value evaluation after the model training is completed.

[0089] In some embodiments, the mean square error is used as the loss function In the formula, N represents the number of training set samples, y t Represents the target reward function value at time t, which is an empirical value. The stochastic gradient descent method is used to minimize the loss function.

[0090] In some embodiments, the following method can be used to schedule tasks in a data center:

[0091] Idle resource monitoring: monitors the resource usage of all servers in the data center, including monitoring the CPU usage, memory usage, and remaining disk space of high-performance computing servers, management servers, and disk arrays;

[0092] At the same time, in addition to monitoring the status of data center equipment as mentioned above, the thresholds of equipment can also be monitored for resource allocation.

[0093] Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center (including available CPU, memory, and disk resources);

[0094] Request processing: receiving resource requests issued by external service requests after entering the data center. The resource request specifies the size of resources (CPU, memory, and hard disk resources) required by the external service request, that is, the request amount;

[0095] Resource search and allocation: traverse the idle resource table and check whether the current idle resources of each server can meet the needs of the external service request;

[0096] If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated;

[0097] Idle resource table update: recalculate the resource usage of the server and update the idle resource table to reflect the latest idle resource situation.

[0098] In some implementations, after the idle resource table is updated, the temperature of the internal hardware of the server under the current resource occupancy condition is fed back to the air cooling system in real time.

[0099] In some implementations, the idle resources of the server are divided into a plurality of idle resource blocks, and the plurality of idle resource blocks have different sizes;

[0100] In the idle resource table, all idle resource blocks are sorted from small to large according to their sizes;

[0101] This step can form an ordered list of idle resource amounts from small to large, which is convenient for finding and allocating the smallest idle resource block that meets the request, thereby avoiding wasting larger idle resource blocks.

[0102] In the resource search and allocation, a first idle resource block is searched, wherein the first idle resource block is a first idle resource block whose size is greater than or equal to the requested amount and is found by traversing the idle resource table from front to back;

[0103] If the size of the first idle resource block is equal to the requested amount, the entire first idle resource block is fully allocated to the external service requester;

[0104] If the size of the first idle resource block is larger than the requested amount, the portion of the first idle resource block equal to the requested amount will be allocated to the external service requester, and the remaining portion will still be retained as an idle resource block in the idle resource table, and its size will be reduced accordingly, that is, equal to the original size minus the requested amount.

[0105] The embodiment of the present application also provides a control system for a data center, including an intelligent data analysis module;

[0106] The intelligent data analysis module is used to obtain environmental data of the data center collected by various sensors in multiple locations of the data center; the obtained environmental data is input into a constructed neural network-based fan speed control model, and the fan speed control model outputs a target speed of each fan in the data center, which is used to control the speed of each fan in the data center; wherein the constructed neural network-based fan speed control model has the function of outputting the target speed of each fan in the data center according to the input environmental data.

[0107] In a possible implementation, the system further includes a task scheduling module for task scheduling, including:

[0108] Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center (including available CPU, memory, and disk resources);

[0109] Request processing: receiving resource requests issued by external service requests after entering the data center. The resource request specifies the size of resources (CPU, memory, and disk resources) required by the external service request, i.e., the request amount;

[0110] Resource search and allocation: traverse the idle resource table and check whether the current idle resources of each server can meet the needs of the external service request;

[0111] If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated;

[0112] Idle resource table update: recalculate the resource usage of the server and update the idle resource table to reflect the latest idle resource situation.

[0113] like Figure 2 As shown in the figure, the task scheduling module monitors the high-performance computing server, the management server, and the disk array respectively, and dynamically allocates the idle resources fed back by them to new external task requests.

[0114] When an external task requests new system resources, the system will select the server that is most suitable for resource allocation based on the above-mentioned task scheduling algorithm. After assigning the task to the server, the system will re-sort the idle resources of the server and feedback the current server chassis temperature and fan speed to the intelligent data analysis module through the task scheduling module to achieve feedback control.

[0115] The intelligent data analysis module optimizes the current fan speed by combining the data center environmental data collected by the current environmental sensor and the temperature and fan speed of the current server chassis fed back by the task scheduling module.

[0116] The solution disclosed in the above embodiment of the present application can realize intelligent environmental management and dynamic task allocation of the data center, can reasonably adjust the data center air cooling system according to the environmental changes in different locations of the data center, and effectively allocate the idle resources of the data center server to improve the working stability of the data center server. It can effectively solve the problems of low environmental heat dissipation efficiency and unreasonable server resource scheduling faced by the current data center, so as to improve the operating stability of the data center server.

[0117] The embodiment of the present application also provides an electronic device, including: a memory and a processor;

[0118] The memory is used to store computer programs;

[0119] The processor is used to call the computer program to execute the method as described above.

[0120] The electronic device also includes a plurality of sensors arranged at multiple locations of the data center for collecting and storing environmental data of the data center.

[0121] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed on an electronic device, the electronic device implements the method as described above.

[0122] An embodiment of the present application further provides a computer program product, including a computer program. When the computer program is executed on an electronic device, the electronic device implements the method as described above.

[0123] The embodiments of the present application also provide a system, an electronic device, a computer-readable storage medium, and a computer program product. The specific implementation methods can refer to the specific embodiments of the above methods and will not be repeated here.

[0124] Obviously, those skilled in the art should understand that the above-mentioned units or steps of the present application can be implemented by a general-purpose computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the present application is not limited to any specific combination of hardware and software.

[0125] The above description of the embodiments of the present application is only a partial embodiment of the present application, which is used to enable professionals in the field to implement or use the content of the present application, and is not used to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A control method for a data center, characterized in that: This includes controlling the speed of individual fans in the data center through the following methods: Obtain environmental data of the data center collected by multiple sensors at multiple locations in the data center; The acquired environmental data is input into a constructed neural network-based fan speed control model, and the fan speed control model outputs a target speed of each fan in the data center, which is used to control the speed of each fan in the data center; wherein the constructed neural network-based fan speed control model has the function of outputting the target speed of each fan in the data center according to the input environmental data.

2. The method according to claim 1, characterized in that Before inputting the collected environmental data into the constructed neural network-based fan speed control model, the method further includes: Before inputting the acquired environmental data into the constructed neural network-based fan speed control model, it also includes: The collected environmental data are preprocessed, including using activation functions to extract the nonlinear features of the environmental data and normalizing each environmental data time series.

3. The method according to claim 2, characterized in that The fan speed control model includes an intramodal attention layer, a cross-modal attention layer, and a fully connected layer: The intramodal attention layer is used to learn the time series information within each modality. The calculation method includes: For the i-th environmental data X i ∈R 1*d , calculate its internal modal attention score matrix A i =softmax(W ai X i +b ai ), where d is the environmental data X i The number of acquisition time points, W ai ∈R d*1 is the weight matrix of linear transformation, b ai is the bias term; according to the result of the calculation of the intramodal attention score matrix, the intramodal weighted feature representation of the i-th environment data is obtained: Y i =A i W ci , where W ci ∈R d*1 is the transformation matrix; The cross-modal attention layer is used to learn the relationship between different modalities. The calculation method includes: For two different environment data X i ∈R 1*d , X j ∈R 1*d , calculate the score Among them, A ij ∈R d*d is the attention weight matrix of the j-th environment data to the i-th environment data, is the scaling factor; According to the calculation results of the attention weight matrix, the cross-modal weighted feature representation of the i-th environment data is obtained: in, n is the number of modes, that is, the number of types of environmental data; The fully connected layer is used to output the fan target speed prediction result based on the learning results of the intramodal attention layer and the cross-modal attention layer. The calculation method includes: Set Y1, Y2, ..., Y n Merge to form Y, Y∈R d*n ; Will Merge, constitute According to X, Y, Get feature representation for decision making Where X is the environmental data matrix, and its i-th row is the environmental data X i , X∈R n*d ; Calculate the decision variable M = (ZV) T t, where V∈R n*m , t∈R d*1 , M∈R m*1 , m is the number of fans in the data center, and M is used to represent the rotation speed of each fan in the data center.

4. The method according to any one of claims 1 to 3, characterized in that The method also includes using the following method to schedule tasks for the data center: Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center; Request processing: receiving resource requests issued by external service requests after entering the data center, wherein the resource requests specify the size of resources required by the external service requests, i.e., the request amount; Resource search and allocation: traverse the idle resource table and check in turn whether the current idle resources of each server can meet the needs of the external service request; If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated; Update the idle resource table: recalculate the resource usage of each server and update the idle resource table.

5. The method according to claim 4, characterized in that The idle resources of the server are divided into a plurality of idle resource blocks, and the sizes of the plurality of idle resource blocks are different; In the idle resource table, all idle resource blocks are sorted from small to large according to their sizes; In the resource searching and allocating, searching for a first idle resource block; The first idle resource block is the first idle resource block whose size is greater than or equal to the requested amount found by traversing the idle resource table from front to back; If the size of the first idle resource block is equal to the requested amount, the entire first idle resource block is fully allocated to the external service requester; If the size of the first idle resource block is larger than the requested amount, the portion of the first idle resource block equal to the requested amount will be allocated to the external service requester, and the remaining portion will still be retained as an idle resource block in the idle resource table, and its size will be reduced accordingly, that is, equal to the original size minus the requested amount.

6. A control system for a data center, characterized in that: It includes an intelligent data analysis module for acquiring environmental data of the data center collected by various sensors at multiple locations of the data center; The acquired environmental data is input into a constructed neural network-based fan speed control model, and the fan speed control model outputs a target speed of each fan in the data center, which is used to control the speed of each fan in the data center; wherein the constructed neural network-based fan speed control model has the function of outputting the target speed of each fan in the data center according to the input environmental data.

7. The system according to claim, characterized in that It also includes a task scheduling module for task scheduling, including: Idle resource table maintenance: Maintain an idle resource table to record the current idle resources of all servers in the data center; Request processing: receiving resource requests issued by external service requests after entering the data center, wherein the resource requests specify the size of resources required by the external service requests, i.e., the request amount; Resource search and allocation: traverse the idle resource table and check in turn whether the current idle resources of each server can meet the needs of the external service request; If a server is found whose idle resource size is greater than or equal to the requested amount, the idle resources in the server are allocated to the external service requester, and the allocated idle resources are marked as allocated; Idle resource table update: recalculate the resource usage of the server and update the idle resource table to reflect the latest idle resource situation.

8. An electronic device, characterized in that: include: Memory and processor; The memory is used to store computer programs; The processor is configured to call the computer program to execute the method according to any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed on an electronic device, the electronic device implements the method according to any one of claims 1 to 4.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed on an electronic device, the electronic device implements the method according to any one of claims 1 to 4.

Images