Server and thermal management system, method, product, equipment and medium thereof

Abstract

The invention discloses a server and a thermal management system and method thereof, a product, equipment and a medium, and relates to the technical field of server management, the system comprises a sensing component, a substrate management controller and an editable controller; the sensing component and the substrate management controller are correspondingly configured at the computing node, the sensing component is connected with the substrate management controller, and the substrate management controller is further connected with the editable controller; the sensing component is configured to collect working parameters of the computing nodes and transmit the working parameters to the substrate management controller; the baseboard management controller is configured to generate a to-be-processed data packet based on the working parameters and transmit the to-be-processed data packet to the editable controller, and the to-be-processed data packet comprises historical working parameters; and the programmable controller is configured to predict a current thermal management strategy for the at least one heat dissipation component by using a prediction model based on the historical working parameters so as to control the at least one heat dissipation component to dissipate heat for the current working state of the corresponding computing node, so that the problem of serious hysteresis is solved, and the effect of improving the hysteresis is achieved.

Description

Technical Field

[0001] This application relates to the technical field of server management, and in particular, to a server and its thermal management system, method, product, device, and medium. Background Art

[0002] During the operation of a server, heat is an inevitable by-product. If not controlled, it may cause the device to overheat, performance degradation, and even failure. In response to this, the server is configured with a thermal management system to control and manage the heat generated by the server to ensure the stable and reliable operation of the server.

[0003] In related technologies, the thermal management system of a server usually follows a control strategy of collecting real-time temperature, viewing temperature changes, and regulating heat dissipation components, resulting in serious hysteresis in thermal management. Summary of the Invention

[0004] This application provides a server and its thermal management system, method, product, device, and medium to at least solve the problem of hysteresis in thermal management in related technologies.

[0005] This application provides a thermal management system for a server. The server includes at least one computing node and at least one heat dissipation component configured on the computing node; the thermal management system includes a sensing component, a baseboard management controller, and an editable controller; The sensing component and the baseboard management controller are correspondingly configured on the computing node. The sensing component is connected to the baseboard management controller, and the baseboard management controller is also connected to the editable controller; The sensing component is configured to collect the working parameters of the computing node and transmit them to the baseboard management controller; The baseboard management controller is configured to generate a data packet to be processed based on the working parameters and transmit it to the editable controller. The data packet to be processed includes historical working parameters; The editable controller is configured to predict the current thermal management strategy for the at least one heat dissipation component by using a prediction model based on the historical working parameters; the prediction model is configured to obtain an output parameter based on an input parameter associated with the historical working parameters, the output parameter is associated with the current thermal management strategy, and the current thermal management strategy is configured to control at least one of the heat dissipation components to dissipate heat for the current working state of the corresponding computing node.

[0006] This application also provides a server, including: the thermal management system of any of the above servers.

[0007] This application also provides a thermal management method for a server, implemented based on the thermal management system of any of the above servers; the thermal management method for the server includes: Obtain the working parameters of the computing node; Generate a data packet to be processed based on the working parameters; wherein, the data packet to be processed includes historical working parameters; Predict the current thermal management strategy for the at least one heat dissipation component based on the historical working parameters by using a prediction model; wherein, the prediction model is configured to obtain an output parameter based on an input parameter associated with the historical working parameters, the output parameter is associated with the current thermal management strategy, and the current thermal management strategy is configured to control at least one of the heat dissipation components to dissipate heat for the current working state of the corresponding computing node.

[0008] This application also provides an electronic device, including: a memory for storing a computer program; a processor for implementing the steps of the thermal management method of any one of the above servers when executing the computer program.

[0009] This application also provides a computer-readable storage medium storing a computer program, wherein the computer program implements the steps of the thermal management method of any one of the above servers when executed by a processor.

[0010] This application also provides a computer program product including a computer program, and the computer program implements the steps of the thermal management method of any one of the above servers when executed by a processor.

[0011] Through this application, since a sensing component is configured for the computing node, a baseboard management controller is connected between the sensing component and the programmable controller, the working parameters of the corresponding computing node are collected based on the sensing component, and a data packet to be processed including historical working parameters is generated based on the baseboard management controller according to the working parameters, and the thermal management strategy of at least one heat dissipation component configured for at least one computing node is predicted based on the historical working parameters by using the programmable controller, so as to control at least one heat dissipation component to dissipate heat for the current working state of the corresponding computing node, the prediction model is configured to obtain an output parameter based on an input parameter associated with the historical working parameters, and the output parameter is associated with the current thermal management strategy, thereby realizing the advance prediction of the current thermal management strategy by using the prediction model in combination with historical working parameters; thus, the lag problem brought by the solution of collecting real-time temperature first and then adjusting in the related art is solved, and the technical effect of reducing thermal management delay is achieved. Description of the Drawings

[0012] To more clearly illustrate the embodiments of the present application, the following will briefly introduce the drawings required for the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0013] Figure 1 A structural schematic diagram of a thermal management system for a server provided by an embodiment of the present application; Figure 2 A structural schematic diagram of another thermal management system for a server provided by an embodiment of the present application; Figure 3 A structural schematic diagram of yet another thermal management system for a server provided by an embodiment of the present application; Figure 4 A structural schematic diagram of yet another thermal management system for a server provided by an embodiment of the present application; Figure 5 A structural schematic diagram of yet another thermal management system for a server provided by an embodiment of the present application; Figure 6 A flowchart of a thermal management method for a server provided by an embodiment of the present application; Figure 7 A flowchart of another thermal management method for a server provided by an embodiment of the present application; Figure 8 A structural schematic diagram of a thermal management device for a server provided by an embodiment of the present application. Detailed implementation manners

[0014] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present application.

[0015] It should be noted that in the description of the present application, the terms "include", "comprise" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or device including a series of elements includes not only those elements but also other elements not expressly listed, or further includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in the present application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0016] To enable those skilled in the art of the present technology to better understand the solution of the present application, the present application will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0017] First, the abbreviations, English full names and corresponding Chinese full names of the technical terms in the present application are described as follows.

[0018] FPGA: Field Programmable Gate Array, Field Programmable Gate Array.

[0019] CPLD: Complex Programmable Logic Device, Complex Programmable Logic Device.

[0020] BMC: Baseboard Management Controller, Baseboard Management Controller, which is a controller used for server management on the server motherboard.

[0021] LSTM: Long Short-Term Memory, Long Short-Term Memory Network.

[0022] FNN: Feedforward Neural Network, Feedforward Neural Network.

[0023] PID: Proportional-Integral-Derivative Control, Proportional-Integral-Derivative Control.

[0024] CPU, Central Processing Unit, Central Processing Unit.

[0025] GPU, Graphic Processing Unit, Graphic Processing Unit.

[0026] RAM, Random Access Memory, Random Access Memory.

[0027] SRAM, Static Random-Access Memory, Static Random-Access Memory.

[0028] I2C: Inter-Integrated Circut, Inter-Integrated Circuit Bus, which is a serial communication bus.

[0029] SQP, Sequential Quadratic Programming, Sequential Quadratic Programming.

[0030] LSTM-FNN Hybrid Model: A composite architecture that combines the Long Short-Term Memory Network (LSTM) and the Feedforward Neural Network (FNN), aiming to capture both the dynamic features of time-series data and the static features of non-time-series data simultaneously. This hybrid model achieves efficient inference through a dynamic data caching mechanism and is applicable to scenarios that require real-time processing of mixed time-series and non-time-series data.

[0031] Secondly, the relevant technical content is described.

[0032] With the rapid development of cloud computing, artificial intelligence, and high-performance computing technologies, multi-node server architectures have become the core infrastructure of data centers. Systems based on multi-node server architectures can integrate 2 to 8 computing nodes in a single 2U-high chassis to collaboratively process large-scale tasks and achieve high-density computing. Due to the increasing demand for high-density servers, the issue of saving server cooling power by controlling fan speed has attracted wide attention.

[0033] In related technologies, multiple computing nodes are integrated in the same chassis, and a shared fan design is adopted. All computing nodes generally use a unified speed regulation strategy, and the fan speed is adjusted by a central controller (such as a CPLD) based on the average temperature or the high value of a single node threshold. Since the fan speed is unified and the air volume demand cannot be predicted in advance, this shared fan control technology can no longer meet the requirements of the next-generation data centers in terms of dynamic adaptability, reliability, and energy efficiency. There is an urgent need for an innovative solution that integrates edge intelligence.

[0034] Specifically, at the hardware architecture level, mainstream multi-node server architectures use the BMC of each node to collect temperature sensor data at various locations and directly send the fan speed data to the control board CPLD through a bus. The control board CPLD uniformly controls the fan speed of all fans and runs the fans at the maximum speed sent by the BMC of each selected node. The unified speed regulation strategy of this multi-node server ignores the load differences between different computing nodes. High-load nodes may have insufficient heat dissipation because the average temperature does not reach the threshold, while low-load nodes waste energy due to the global speed increase. There is a lack of cross-node cooperation mechanism and the inability to dynamically allocate heat dissipation resources according to the load, resulting in coexistence of insufficient heat dissipation in high-load nodes and excessive heat dissipation in low-load nodes.

[0035] At the same time, at the control logic level, static threshold / fixed threshold control (such as when the CPU temperature > 85°C, the fan runs at full speed), regression functions, and other fixed function algorithms are generally used, without considering the heterogeneity and dynamic changes of multiple heat sources (CPU / GPU / hard disk) inside different computing nodes. Specifically, using fixed threshold or fixed function strategies cannot predict and respond to the rapid temperature rise of multiple heat sources inside the node in real time (such as a local temperature gradient > 15°C / s caused by a sudden GPU load). The dynamic load matching ability is seriously insufficient, and it is easy to cause local overheating or power consumption waste. In addition, the thermal management systems of some multi-node server architectures use the PID algorithm to achieve local closed-loop control and cannot achieve global overall planning.

[0036] In addition, the current control technologies for fans ultimately follow a control strategy of collecting real-time temperatures, checking temperature changes, and increasing or decreasing fan speeds. This control strategy has serious hysteresis, and when the temperature rises rapidly, it is easy for key components such as CPUs and GPUs to experience high temperatures for a period of time, causing damage to the key components. At the same time, it will also increase unnecessary power consumption when the temperature drops.

[0037] Generally speaking, the related technical solutions have systematic defects in terms of dynamic adaptability, reliability, and energy efficiency, and it is difficult to support the stringent requirements of the next-generation data centers for efficient heat dissipation and high-reliability operation. There is an urgent need to achieve breakthroughs through architectural innovation and control method upgrades.

[0038] To address at least one of the above technical problems, an embodiment of the present application proposes a thermal management system mainly for multi-node servers. The multi-node server may include multiple nodes and multiple fans. The thermal management system may also be applicable to servers with a single node and multiple fans, or servers with multiple nodes and a single fan, without limitation here. The thermal management system may include a sensing component and a baseboard management controller configured for different computing nodes, and an editable controller connected to the baseboard management controller. Among them, the sensing component can collect the working parameters of the corresponding computing node, and based on the baseboard management controller, generate a data packet to be processed including historical working parameters, and based on the editable controller, use a prediction model to predict the thermal management strategy of at least one heat dissipation component configured for at least one computing node according to the historical working parameters. The prediction model can obtain output parameters based on input parameters associated with the historical working parameters, and the output parameters are associated with the current thermal management strategy. The current thermal management strategy is configured to control the operation of at least one heat dissipation component configured for at least one computing node, so as to control at least one heat dissipation component to dissipate heat for the current working state of the corresponding computing node, thereby realizing the advance prediction of the current thermal management strategy by combining historical working parameters, thereby improving the dynamic adaptability of the thermal management system, improving thermal management latency, thereby improving the damage to key components, and reducing unnecessary power consumption.

[0039] The embodiment of the present application can implement a thermal management system based on an editable device, such as FPGA edge intelligence. Specifically, for a server with a fan as the heat dissipation component, by deploying a current thermal management strategy prediction model module on the editable device, the current thermal management strategy prediction model module deploys the prediction model. The baseboard management controllers of different computing nodes send data packets to be processed to the editable controller. The editable device combines the prediction model configured by the current thermal management strategy prediction model module to calculate a global optimal rotational speed allocation scheme, suppress priority contention conflicts, generate a three-dimensional heat dissipation demand vector (target air volume demand, priority weight, and thermal inertia coefficient), and finally combine the constraint conditions to achieve optimal fan rotational speed control output, thereby realizing high-precision and low-latency distributed heat dissipation control.

[0040] Among them, aiming at the load difference between different computing nodes, targeted heat dissipation control can be realized based on global distributed management, cross-node collaborative control can be achieved, insufficient heat dissipation of high-load nodes can be improved, and excessive heat dissipation of low-load nodes can be improved.

[0041] In addition, the editable device can read back the actual rotation speed and the expected rotation speed of the fan to the baseboard management controller and display them based on a device with a display function under the baseboard management controller network. And this solution has the advantages of accurate temperature control, energy efficiency improvement, and operation and maintenance cost reduction.

[0042] The embodiments of the present application will be described below with reference to the accompanying drawings.

[0043] Exemplarily, Figure 1 is a schematic structural diagram of a thermal management system of a server provided by an embodiment of the present application. Refer to Figure 1 , the thermal management system of the server is used for thermal management of the server, such as dissipating heat from heat sources in the server. Exemplarily, the server may include at least one computing node 01 and at least one heat dissipation component 02 configured on the computing node 01. Exemplarily, the server includes one computing node 01 and one or more heat dissipation components 02 for dissipating heat from the computing node 01; or, the server may include multiple computing nodes 01 and one or more heat dissipation components 02 for dissipating heat from the multiple computing nodes 01; or, the server may include multiple computing nodes 01 and one or more heat dissipation components 02 for dissipating heat from each computing node 01. Refer to Figure 1 .

[0044] Among them, the computing node 01 is a server node with independent computing capabilities and can execute specific tasks, and may include a motherboard and devices such as a CPU, a GPU, and a hard disk configured on the motherboard. The heat dissipation component 02 is a component for dissipating heat from the computing node 01 and may include components such as a fan.

[0045] Continuing to refer to Figure 1 , the thermal management system 100 of the server includes a sensing component 10, a baseboard management controller 11, and an editable controller 12; and the sensing component 10 and the baseboard management controller 11 are correspondingly configured on the computing node 01, the sensing component 10 is connected to the baseboard management controller 11, and the baseboard management controller 11 is further connected to the editable controller 12. Specifically, the baseboard management controller 11 is configured on the motherboard and is connected to the sensing component 10 configured on the motherboard to transmit the data associated with the computing node 01 to the editable controller 12.

[0046] Among them, the sensing component 10 is configured to collect the working parameters of the computing node 01 and transmit them to the baseboard management controller 11. Specifically, the working parameters may be parameters related to thermal management. Exemplarily, the working parameters may include temperature and power consumption, as well as other parameters affecting heat generation, which are not limited herein. The working parameters collected by the sensing component 10 are transmitted to the baseboard management controller 11 connected to the sensing component 10, and are integrated by the baseboard management controller 11 and then transmitted to the programmable controller 12.

[0047] Among them, the baseboard management controller 11 is configured to generate a data packet to be processed based on the working parameters and transmit it to the programmable controller 12. The data packet to be processed includes historical working parameters, so that the programmable controller 12 can identify the change trend of the working state of the corresponding computing node 01 based on at least the historical working parameters, and predict the current thermal management strategy for the current working state accordingly.

[0048] Specifically, the baseboard management controller 11 can perform data integration according to the working parameters to generate a data packet to be processed. Exemplarily, the data packet to be processed may include data directly collected by the sensing component 10 such as temperature and power consumption, or may include calculated data such as temperature gradient and power consumption gradient obtained by performing data calculations based on temperature and power consumption, which are not limited herein.

[0049] Among them, the programmable controller 12 can be arranged on the control board and is configured to predict the current thermal management strategy for at least one heat dissipation component 02 based on the historical working parameters; the prediction model is configured to obtain an output parameter based on the input parameter associated with the historical working parameters, and the output parameter is associated with the current thermal management strategy. The current thermal management strategy is configured to control at least one heat dissipation component 02 to dissipate heat for the current working state of the corresponding computing node 01.

[0050] Specifically, the programmable controller 12 can use the prediction model to identify the change trend of the working state of the corresponding different computing nodes 01 based on at least the historical working parameters, and predict the current thermal management strategy matching the current working state of the computing nodes accordingly, so as to control at least one heat dissipation component 02 configured for at least one computing node 01 based on the current thermal management strategy to dissipate heat for the current working state of the computing node 01 and improve the thermal management delay.

[0051] In the embodiment of the present application, the sensing component 10 is a component deployed in the sensing layer and is responsible for receiving the working parameters of the server node, that is, the sensing layer information; the sensing layer information is collected by the BMC through the bus. The BCM and the programmable controller 12 are components deployed in the control layer and are responsible for processing the sensing layer information to obtain the current thermal management strategy. The heat dissipation component 02 is a component deployed in the execution layer and can be controlled by the current thermal management strategy predicted by the programmable controller 12 based on the prediction model to achieve global distributed heat dissipation control.

[0052] Specifically, in the embodiments of the present application, a sensing component 10 is configured for the computing node 01, and a baseboard management controller 11 is connected between the sensing component 10 and the programmable controller 12. The working parameters of the corresponding computing node 01 are collected based on the sensing component 10, and a data packet to be processed including historical working parameters is generated based on the baseboard management controller 11 according to the working parameters. Moreover, based on the programmable controller 12, a thermal management strategy for at least one heat dissipation component 02 configured for at least one computing node 01 is predicted using a prediction model according to the historical working parameters, so as to control at least one heat dissipation component 02 to dissipate heat for the current working state of the corresponding computing node 01, thereby realizing the advance prediction of the current thermal management strategy in combination with at least historical working parameters and improving the thermal management delay.

[0053] In some embodiments, Figure 2 FIG. is a schematic structural diagram of another thermal management system of a server provided by an embodiment of the present application. Based on Figure 1 and with reference to Figure 2 , in the thermal management system 100 of the server, the programmable controller 12 at least includes: a data preprocessing module 121 and a current thermal management strategy prediction model module 120; the current thermal management strategy prediction model module 120 is connected to the data preprocessing module 121.

[0054] Among them, the data preprocessing module 121 is configured to generate input parameters based on historical working parameters and transmit them to the current thermal management strategy prediction model module 120. The current thermal management strategy prediction model module 120 is configured to perform parameter prediction based on the input parameters to obtain output parameters, and the output parameters are associated with the current thermal management strategy.

[0055] Specifically, the data preprocessing module 121 can directly or indirectly receive historical working parameters, generate input parameters that match the current thermal management strategy prediction model module 120 based on at least the historical working parameters, and transmit the input parameters to the current thermal management strategy prediction model module 120. It can be understood that the data preprocessing module 121 is configured in the programmable controller 12 to reduce the data processing amount on the baseboard management controller 11. In other embodiments, the data preprocessing module 121 can also be configured in the baseboard management controller 11 to perform partial processing operations on the baseboard management controller 11, which is not limited herein.

[0056] Among them, the current thermal management strategy prediction model module 120 can configure a prediction model, and the prediction model can be a network model or a network hybrid model, so as to identify the change trend of the working state of the computing node based on at least historical working parameters and output output parameters associated with the current thermal management strategy. Its specific association relationship will be exemplarily described later.

[0057] In an embodiment of the present application, by setting the editable controller 12 to include a data preprocessing module 121 and a current thermal management strategy prediction model module 120, the data preprocessing module 121 can be used to calculate working parameters to obtain input parameters that meet the input requirements of the current thermal management strategy prediction model module 120. Correspondingly, the prediction model of the current thermal management strategy prediction model module 120 can output output parameters related to the current thermal management strategy according to the input parameters, so as to realize the prediction of the current thermal management strategy.

[0058] In some embodiments, Figure 3 is a schematic structural diagram of another thermal management system for a server provided by an embodiment of the present application. On the Figure 1 basis, referring to Figure 3 , in the thermal management system 100 of the server, the heat dissipation component 02 may include a fan. Exemplarily, when the number of computing nodes is n, the execution layer may be composed of n fans, and each computing node may be configured with 1-2 three-phase brushless DC fans, and the fans can support stepless speed regulation from 0 to 100% through PWM signals to achieve precise speed control. In other embodiments, the heat dissipation component 02 may further include other structural components with heat dissipation functions, which will not be elaborated or limited herein.

[0059] In some embodiments, continuing to refer to Figure 3 , in the thermal management system 100 of the server, the sensing component 10 includes at least a temperature sensing component 101 and a power consumption sensing component 102, and the working parameters of the computing node 01 include at least temperature and power consumption. Specifically, the temperature sensing component 101 may include a temperature sensor for collecting the temperature of the corresponding computing node 01. Exemplarily, the temperature sensor may adopt EMC1413. The power consumption sensing component 102 may include a power consumption sensor for collecting the power consumption of the corresponding computing node.

[0060] Based on this, the data packet to be processed includes at least the temperature and power consumption at the current moment and multiple historical moments before. Therefore, the data calculation process is mainly executed in the editable controller 12, reducing the data processing volume in the baseboard management controller 11. The input parameters include the current temperature gradient, power change rate, dynamic weight, and multiple temperature gradients and multiple power gradients within a historical preset duration, so as to identify the change trend of the working state of the computing node; the output parameters include the target air volume demand, priority weight, and thermal inertia coefficient, so as to further process to obtain the current thermal management strategy. The current thermal management strategy includes the current air volume demand and / or current speed demand of at least one heat dissipation component 02, so as to realize the regulation of the heat dissipation component, such as the fan, to match the current working state of the computing node, and realize precise heat dissipation control.

[0061] In the embodiments of the present application, the sensing layer where temperature sensors and power consumption sensors are deployed can sense the temperature information and power consumption information of the corresponding computing node 01, and upload them to the programmable controller 12 by the baseboard management controller 11 of the corresponding node; the control layer where the baseboard management controller 11 and the programmable controller 12 are deployed is mainly responsible for the temperature information and power consumption information transmitted by the sensing layer, monitoring the current rotation speed of the fans in the execution layer, and outputting the optimal rotation speed associated with the current thermal management strategy to the execution layer, so as to achieve the global distribution control of the fans, improve the latency, and adapt to the load differences of different computing nodes.

[0062] In some embodiments, in the thermal management system 100 of the server, the current thermal management strategy prediction model module 120 configures a hybrid model of a long short-term memory network and a feedforward neural network (LSTM-FNN); the long short-term memory network is configured to output a hidden state based on multiple temperature gradients and multiple power gradients within a preset historical duration in the input parameters; the long short-term memory network is configured to determine the target air volume demand, priority weight, and thermal inertia coefficient based on the hidden state and the current temperature gradient, power change rate, and dynamic weight in the input parameters.

[0063] Specifically, the current thermal management strategy prediction model module 120 is the main module for predicting the current thermal management strategy, which can receive the input parameters obtained after preprocessing and output the output parameters associated with the current thermal management strategy.

[0064] Specifically, the LSTM-FNN hybrid model is a composite architecture model that combines a long short-term memory network (LSTM) and a fully connected neural network (FNN), which can capture both the dynamic characteristics of time series data and the static characteristics of non-time series data. In other embodiments, other large model types can also be used to predict the target required air volume, which is not limited here.

[0065] In the embodiments of the present application, the input parameters of the LSTM-FNN hybrid model include time series data and non-time series data; among them, the time series data may include: the current temperature gradient , the power change rate and the dynamic weight ; the non-time series data may include historical sequences, such as the temperature gradient in the past 5 seconds and the power consumption gradient .

[0066] In response to this, the data preprocessing module 121 correspondingly completes the calculation of the current temperature gradient , the power change rate and the dynamic weight for input into the LSTM-FNN hybrid model.

[0067] Exemplarily, the current temperature gradient It can be obtained by calculating the average value based on multiple temperature gradients. For example, it can be obtained by calculating the neighborhood temperature gradient with the transmission temperature data of the previous two cycles. The specific calculation formula is as follows:

[0068] Wherein, represents the current temperature gradient, represents the temperature gradient at the (i - 1)th moment, represents the temperature gradient at the ith moment, represents the temperature gradient at the (i + 1)th moment; i represents the moment.

[0069] Exemplarily, the power change rate can be obtained by calculating the second-order difference of power based on the power sequence in the past 3 seconds. The specific calculation formula is as follows:

[0070] Wherein, represents the power change rate, represents the power at the ith moment, represents the power at the (i - 1)th moment, represents the power at the (i - 2)th moment; i represents the moment.

[0071] Exemplarily, the dynamic weight, which can also be called the dynamic decay factor, is used to balance short-term power fluctuations and long-term load trends. The specific calculation formula is as follows:

[0072] Wherein, represents the dynamic weight, that is, the dynamic decay factor, represents the reference weight, and t represents the tth time step; represents the decay coefficient, and its value can be 0.05, represents the acceleration weight, and its value can be 0.3. This parameter can be adjusted according to the actual situation and is not limited here.

[0073] Thus, based on the data preprocessing module 121, the current temperature gradient , the power change rate and the dynamic weight can be calculated and input into the LSTM-FNN hybrid model.

[0074] Exemplarily, in this LSTM-FNN hybrid model, the LSTM layer can include 128 neurons and is a bidirectional structure: it inputs the historical sequence, processes the long-term correlation of power changes (such as GPU burst load prediction); the final state of the LSTM in the previous window is used as the initial state of the next window to maintain the temporal continuity; it outputs the hidden state, specifically as follows:

[0075] Among them, the hidden state is the context-aware representation generated by the LSTM after processing the input at each time step, which synthesizes the current input and the hidden state information of the previous moment. Generally, represents the hidden state at the t-th time step, and its calculation depends on the current input , the hidden state of the previous moment and the cell state .

[0076] Among them, the FNN layer in the LSTM-FNN hybrid model receives the output of the LSTM layer and real-time features, and calculates the non-linear mapping, specifically as follows:

[0077] Among them, is the target air volume demand (CFM), is the final priority weight, that is, the bidding value in the following text; is the thermal inertia coefficient.

[0078] Thus, the required air volume predicted based on the LSTM-FNN hybrid model can be obtained. After further processing, such as constraint checking and multi-node air volume distribution, the final current thermal management strategy can be obtained.

[0079] In the embodiments of the present application, by deploying the LSTM-FNN hybrid model on the programmable controller 12, such as FPGA, the baseboard management controller 11 configured on each computing node 01 sends the data packet to be processed to the programmable controller 12 through the bus. The programmable controller 12 outputs parameters based on the LSTM-FNN hybrid model to predict the current thermal management strategy, specifically, it can calculate the global optimal rotation speed allocation scheme, suppress the priority contention conflict, generate a three-dimensional heat dissipation demand vector, and finally realize the optimal fan speed control output to improve the dynamic load adaptability of the multi-node server thermal management system and improve the coordination efficiency of the distributed heat dissipation method.

[0080] In some embodiments, Figure 4 is a schematic structural diagram of another thermal management system of the server provided by the embodiments of the present application. On the basis of Figure 1 , referring to Figure 4 , in the thermal management system 100 of the server, the programmable controller 12 may further include: an information transceiver and conversion module 122 and a dynamic data cache module 123; the dynamic data cache module 123 is connected between the information transceiver and conversion module 122 and the data preprocessing module 121.

[0081] Among them, the information transceiver and conversion module 122 is configured to receive a data packet to be processed, parse the data packet to obtain target data, and transmit the target data to the dynamic data cache module 123.

[0082] Specifically, the programmable controller 12 and the baseboard management controller 11 can be connected based on the I2C bus or other buses, and data transmission can be realized based on the bus. Moreover, the baseboard management controller 11 is connected to the information transceiver and conversion module 122 of the programmable controller 12, and the information transceiver and conversion module 122 can parse the bus data content of the baseboard management controller 11. In the embodiments of the present application, the baseboard management controller 11 mainly transmits a data packet to be processed of the current computing node to the programmable controller 12. The data packet to be processed may include the temperature and power consumption of the current computing node, or may include data obtained after data calculations such as the average temperature, average power consumption, and priority weight of the current computing node, which is not limited herein.

[0083] Among them, the dynamic data cache module 123 is configured to temporarily store and transmit the target data to the data preprocessing module 121 based on at least a dual-buffer structure.

[0084] In the embodiments of the present application, by setting the dynamic data cache module 123, the target data can be temporarily stored and transmitted based on the dual-buffer structure, thereby solving the problem of mismatch between the data transmission speed and the data processing speed of the programmable controller 12.

[0085] Specifically, when the programmable controller 12 receives the data information transmitted by the baseboard management controller 11, in order to prevent the data processing speed from being less than the signal transmission speed between the programmable controller 12 and the baseboard management controller 11, at least a dual-buffer structure is adopted in the programmable controller 12. For example, a dynamic data cache module 123 with a dual-port RAM architecture stores data, avoiding the loss of subsequent transmission signals caused by too slow data processing speed, so as to improve data accuracy and improve thermal management accuracy. At the same time, by deploying an LSTM-FNN hybrid model on the programmable controller 12, such as an FPGA, the baseboard management controller 11 configured for each computing node 01 sends a data packet to be processed to the programmable controller 12 through the bus, and the programmable controller 12 outputs parameters based on the LSTM-FNN hybrid model to predict the current thermal management strategy. Specifically, the global optimal rotational speed allocation scheme can be calculated, the priority contention conflict can be suppressed, a three-dimensional heat dissipation demand vector can be generated, and finally the optimal fan rotational speed control output can be realized to improve the dynamic load adaptability of the multi-node server thermal management system and improve the coordination efficiency of the distributed heat dissipation method.

[0086] In some embodiments, continue to refer to Figure 4, in the thermal management system 100 of the server, the dynamic data caching module 123 can adopt a dual-buffer structure, specifically including an input data selection unit 1231, a first caching unit 1232, a second caching unit 1233, and a data output selection unit 1234; the input end of the input data selection unit 1231 is connected to the information transceiver and conversion module 122, the output end of the input data selection unit 1231 is connected to the input ends of the first caching unit 1232 and the second caching unit 1233, the output ends of the first caching unit 1232 and the second caching unit 1233 are both connected to the data output selection unit 1234, and the data output selection unit 1234 is also connected to the data preprocessing module 121.

[0087] Among them, the input data selection unit 1231 is configured to alternately direct the target data at intervals to the first caching unit 1232 and the second caching unit 1233; the first caching unit 1232 and the second caching unit 1233 are configured to alternately receive the updated target data at intervals; the data output selection unit 1234 is configured to alternately transmit the updated target data in the first caching unit 1232 and the second caching unit 1233 to the data preprocessing module 121 at intervals.

[0088] Implementing dynamic data caching in this way can prevent the bus data transmission speed from being greater than the data processing speed of the editable controller 12, avoid information loss, improve data accuracy, and improve thermal management accuracy.

[0089] Specifically, the dynamic data caching module 123 adopts a ping-pong transmission mechanism. Ping-pong transmission is a data stream control method based on a dual-buffer mechanism. By alternately using a dual-port SRAM buffer (storage space ≥ 128KB), seamless connection of input and output is achieved, ensuring the continuity of the data stream.

[0090] Exemplarily, the dynamic data caching process may include: in the first buffer cycle, the input data selection unit 1231 directs the data stream to the first buffer unit 1232 for writing. At this time, there is no data in the second buffer unit 1233, and the data output selection unit 1234 remains idle; in the second buffer cycle, the input data selection unit 1231 switches to the second buffer unit 1233 to continue writing new data, and the data output selection unit 1234 reads the data of the previous cycle from the first buffer unit 1232 and transmits it to the current thermal management policy prediction model module 120 for operation, realizing parallel writing and reading operations; in the third buffer cycle, the input data selection unit 1231 switches back to the first buffer unit 1232 to overwrite the old data, realizing the data update in the first buffer unit 1232. The data output selection unit 1234 turns to the second buffer unit 1233 to read the data of the current cycle and transmits it to the current thermal management policy prediction model module 120 for operation. In this way, a continuous alternating cycle of "writing - processing" is formed to realize data staging and transmission. Among them, the length of each buffer cycle needs to match the data processing time, and the switching signal is controlled by the beat to ensure that the input and output are not interrupted.

[0091] In some embodiments, continuing to refer to Figure 4 , in the thermal management system 100 of the server, the programmable controller 12 may further include: a rotation speed signal reading and feedback module 124, and the rotation speed signal reading and feedback module 124 is connected to the information transceiver and conversion module 122.

[0092] Among them, the rotation speed signal reading and feedback module 124 is configured to obtain the actual rotation speed of the fan and transmit it to the information transceiver and conversion module 122.

[0093] Specifically, the heat dissipation component 02, such as a fan, may integrate a Hall sensor or an optical encoder, and the actual rotation speed of the fan is monitored in real time by transmitting the TACH signal back. Among them, TACH is the abbreviation of Tachometer, indicating the feedback signal of the fan rotation speed, usually in the form of pulse frequency (unit: Hz) or rotation speed (unit: RPM). It is generated by the Hall sensor or optical encoder of the fan and reflects the number of rotations of the fan rotor. The programmable controller 12 can calculate the actual rotation speed of the fan through the TACH value transmitted back by the fan in the rotation speed signal reading and feedback module 124. And the obtained actual rotation speed can be transmitted back to the information transceiver and conversion module 122 for interaction with the baseboard management controller 11 or other modules in the programmable controller 12.

[0094] Among them, the information transceiver and conversion module 122 is further configured to forward the actual rotation speed to the baseboard management controller 11. The baseboard management controller 11 is further configured with a display module, and the display module is configured to at least display the actual rotation speed.

[0095] Specifically, the information transceiver and conversion module 122 transmits the actual rotation speeds of at least some or all of the fans to the baseboard management controller 11, and can read back the expected rotation speeds of at least some or all of the fans to the baseboard management controller 11. Correspondingly, the baseboard management controller 11 can use its configured display module to display at least the actual rotation speeds so that relevant personnel can intuitively understand the current operating conditions of the fans.

[0096] In addition, the baseboard management controller 11 can log the actual rotation speeds and the corresponding expected rotation speeds of the fans for traceability.

[0097] In addition, the interaction information between the programmable controller 12 and the baseboard management controller 11 can also include information such as whether to forcibly control the fan rotation speed, the forcibly controlled fan nodes, and the rotation speeds of the controlled fan nodes, so that relevant personnel can intervene for forced thermal management control, realizing the combination of automatic control and manual control, and flexibly meeting the usage requirements of different scenarios. The information transceiver and conversion module 122 of the programmable controller 12 can parse the interaction information (including the data packets to be processed and the information related to forced control) in real time according to the agreed communication format, and transmit the parsed information to the dynamic data cache module 123.

[0098] In some embodiments, the data preprocessing module 121, the information transceiver and conversion module 122, and the dynamic data cache module 123 can be uniformly set as data transceiver, parsing, processing, and sending modules, such as Figure 4 .

[0099] In some embodiments, in the thermal management system 100 of the server, the information transceiver and conversion module 122 is further configured to forward the actual rotation speed to the current thermal management policy prediction model module 120; the current thermal management policy prediction model module 120 is further configured to update the parameters of the configured model based on at least the actual rotation speed.

[0100] Specifically, the current thermal management policy prediction model module 120 configures a prediction model, such as an LSTM-FNN hybrid model. The information transceiver and conversion module 122 can also transmit the processed actual rotation speed to the current thermal management policy prediction model module 120 for model parameter update, so as to realize the training feedback update of the prediction model, improve the model accuracy, improve the prediction accuracy, and thus improve the thermal management accuracy. In addition, the temperature prediction value can be compared with the temperature deviation value in combination with the collected temperature, and the parameters of the prediction model can be updated in real time to further improve the model accuracy, improve the prediction accuracy, and thus improve the thermal management accuracy.

[0101] In some embodiments, Figure 5 is a schematic structural diagram of another thermal management system of a server provided by an embodiment of the present application. Refer to Figure 5, in the thermal management system 100 of the server, the editable controller 12 may further include: a constraint check and multi-node air volume distribution module 125, and the constraint check and multi-node air volume distribution module 125 is connected to the current thermal management strategy prediction model module 120. Figure 5 In the illustrated thermal management system 100, the current thermal management strategy prediction model module may further include a classifier, and the classifier is connected after the FNN layer.

[0102] Among them, the constraint check and multi-node air volume distribution module 125 is configured to determine a competition packet based on a priority weight, a thermal inertia coefficient, and a current temperature gradient; and determine the current air volume demand and / or the current rotation speed demand of the heat dissipation component 02 configured for different computing nodes 01 based on the competition packet, the target air volume demand, and the target constraint conditions; wherein the target constraint conditions include a total air volume limit condition and a fan rotation speed limit condition. The target constraint conditions can be understood as the constraint conditions defined by the current server hardware structure.

[0103] Specifically, the constraint check and multi-node air volume distribution module 125 is used to further process the output parameters of the current thermal management strategy prediction model module 120 to obtain a current thermal management strategy that meets the constraint conditions matching the current server hardware structure.

[0104] Specifically, after obtaining the target air volume demand, multi-node fan air volume distribution balance is performed based on the constraint check and multi-node air volume distribution module 125. Exemplarily, the constraint check and multi-node air volume distribution module 125 may use Nash equilibrium scheduling. The specific processing steps are as follows: The first step, competition packet Generation: Each computing node submits a corrected bid value (i.e., the final priority weight) , thermal inertia coefficient and average temperature (i.e., the current temperature gradient) to participate in the calculation. The specific calculation formula is as follows:

[0105] The second step, perform global optimization.

[0106] Specifically, subject to the actual situation of the server hardware structure, constraint conditions are added. Exemplarily, based on the cabinet limit, a total air volume upper limit is added: , where represents the air volume upper limit value, represents the total air volume; and based on the fan rotation speed limit, a feasible interval of the rotation speed R i is added: .

[0107] The third step, establish an objective function to maximize the total revenue.

[0108]

[0109] The methods for solving the above objective function may specifically include: using sequential quadratic programming to solve the mixed-integer programming problem in real time in the hardware pipeline to obtain the final air volume distribution of the fan for different computing node configurations. In some embodiments, the above modules can all be implemented based on the hardware circuit on the FPGA to implement the data processing process of the programmable controller 12 based on the FPGA hardware circuit.

[0110] The constraint check and multi-node air volume distribution module 125 obtains the air volume distribution of the fans with the optimal configuration of each computing node under the target constraint conditions according to the above steps, that is, obtains the current thermal management strategy. Further, the air volume demand can be converted into the fan speed according to the known curve of the fan and transmitted to the speed selection and rotation control module 128 to generate a PWM waveform based on the fan speed to control the rotation of the fan. For details, refer to the following text.

[0111] In the embodiments of the present application, the required air volume predicted based on the LSTM-FNN hybrid model can be obtained, and after constraint check and multi-node air volume distribution, the final current thermal management strategy can be obtained.

[0112] In the embodiments of the present application, by deploying the LSTM-FNN hybrid model on the FPGA, the baseboard management controller 11 configured for each computing node 01 sends the data packet to be processed to the programmable controller 12 through the bus. The programmable controller 12 outputs parameters based on the LSTM-FNN hybrid model to predict the current thermal management strategy. Specifically, the global optimal speed distribution scheme can be calculated, the priority contention conflict can be suppressed, the three-dimensional heat dissipation demand vector can be generated, and constraint check and multi-node air volume distribution, as well as speed conversion, can be performed. Finally, the optimal fan speed control output can be realized to improve the dynamic load adaptability of the multi-node server thermal management system and improve the coordination efficiency of the distributed heat dissipation method.

[0113] In some embodiments, continue to refer to Figure 5 , in the thermal management system 100 of the server, the programmable controller 12 may further include: a baseboard management controller status monitoring module 126, a programmable controller preset speed output module 127, and a speed selection and rotation control module 128. The baseboard management controller status monitoring module 126, the programmable controller preset speed output module 127, and the constraint check and multi-node air volume distribution module 125 are all connected to the speed selection and rotation control module 128.

[0114] Among them, the baseboard management controller status monitoring module 126 is configured to obtain the working status of the baseboard management controller 11 configured for the computing node 01, and the working status includes normal working and abnormal working.

[0115] Specifically, the embodiment of the present application imports an exception state fault tolerance mechanism based on the baseboard management controller status monitoring module 126. Specifically, in order to prevent the baseboard management controller 11 of some computing nodes from hanging, resulting in the editable controller 12 being unable to obtain the working parameters of the computing node 01 or misjudging, the baseboard management controller status monitoring module 126 is set. Exemplarily, the baseboard management controller status monitoring module 126 can monitor the heartbeat signal of the baseboard management controller 11 according to a predetermined scheme. Generally, within 5 minutes after the AC signal is powered on, the heartbeat signal of the baseboard management controller 11 starts to jump, and the heartbeat stop state within no more than 10s is the normal working state. In other cases, it corresponds to an abnormal heartbeat signal of the baseboard management controller 11. At this time, the rotation speed selection and rotation control module 128 of the editable controller 12 is notified to use the preset rotation speed of the editable controller 12, and this preset rotation speed can be output by the preset rotation speed output module 127 of the editable controller.

[0116] Among them, the preset rotation speed output module 127 of the editable controller is configured to generate a preset rotation speed for fan control according to a preset thermal management strategy.

[0117] Specifically, when the heartbeat signal of the baseboard management controller 11 of the computing node is abnormal, that is, when the working state of the baseboard management controller 11 is an abnormal working state, the rotation speed selection and rotation control module 128 of the editable controller 12 will stop using the allocated air volume output by the constraint check and multi-node air volume distribution module 125, that is, stop using the air volume predicted by the hybrid model; at this time, in order to keep the fan working for a certain degree of heat dissipation, the preset rotation speed output module 127 of the editable controller will generate a fixed or stepped preset rotation speed according to a predetermined thermal management strategy (such as monitoring the startup situation of this node or the in-position strategy of important facilities, etc.), and transmit it to the rotation speed selection and rotation control module 128 for actual fan control.

[0118] Among them, the rotation speed selection and rotation control module 128 is configured to receive the working state of the baseboard management controller 11, and based on the working state being normal, control the fan rotation using the current thermal management strategy; or, based on the working state being abnormal, control the fan rotation using the preset rotation speed.

[0119] Specifically, the rotation speed selection and rotation control module 128 can select the rotation speed according to the operating state of the baseboard management controller 11 given by the baseboard management controller status monitoring module 126. Among them, when the baseboard management controller 11 is operating normally, that is, when the baseboard management controller 11 is healthy, the rotation speed predicted and allocated by the hybrid model is used, that is, the current thermal management strategy is adopted; when the baseboard management controller 11 is operating abnormally, that is, when the baseboard management controller 11 is unhealthy, the preset rotation speed is used, that is, the fixed rotation speed or the stepped rotation speed, so as to still be able to control the fan rotation when the baseboard management controller 11 is abnormal. Exemplarily, the rotation speed can be converted into a PWM waveform in the rotation speed selection and rotation control module 128 and transmitted to the fan group for fan rotation.

[0120] Based on the above embodiments, the embodiment of the present application further provides a server, which may include the thermal management system of any server provided in the above embodiments, and has corresponding beneficial effects. For related content, reference can be made to the foregoing, and details are not described herein again.

[0121] In other embodiments, the server may further include other structural components, which are not limited herein.

[0122] Based on the above embodiments, the embodiment of the present application further provides a thermal management method for a server, which can be implemented based on the thermal management system of any server provided in the above embodiments. For related content, reference can be made to the foregoing, and details are not described herein again.

[0123] Exemplarily, Figure 6 is a schematic flowchart of a thermal management method for a server provided by an embodiment of the present application. Referring to Figure 6 , the thermal management method may include the following steps.

[0124] S21. Obtain the operating parameters of the computing node.

[0125] Exemplarily, the operating parameters of the computing node can be collected based on the sensing components.

[0126] S22. Generate a data packet to be processed based on the operating parameters; wherein, the data packet to be processed includes historical operating parameters.

[0127] Exemplarily, the baseboard management controller can generate a data packet to be processed including historical operating parameters according to the operating parameters.

[0128] S23. Predict the current thermal management strategy for at least one heat dissipation component based on the historical operating parameters by using a prediction model; wherein, the prediction model is configured to obtain output parameters based on input parameters associated with the historical operating parameters, the output parameters are associated with the current thermal management strategy, and the current thermal management strategy is configured to control at least one heat dissipation component to dissipate heat for the current operating state of the corresponding computing node.

[0129] Exemplarily, the current thermal management strategy for at least one heat dissipation component can be predicted based on an editable controller according to historical working parameters.

[0130] In an embodiment of the present application, the working parameters of the corresponding computing node can be collected based on a sensing component, and a data packet to be processed including historical working parameters can be generated based on a baseboard management controller according to the working parameters. Further, a thermal management strategy for at least one heat dissipation component configured for at least one computing node can be predicted based on an editable controller according to the historical working parameters by using a prediction model. The prediction model is configured to obtain an output parameter based on an input parameter associated with the historical working parameters, and the output parameter is associated with the current thermal management strategy, so as to control at least one heat dissipation component to dissipate heat for the current working state of the corresponding computing node, thereby realizing the advance prediction of the current thermal management strategy in combination with at least historical working parameters. Therefore, the hysteresis problem brought by the solution of collecting real-time temperature first and then performing regulation in the related art is solved, and the technical effect of reducing thermal management delay is achieved.

[0131] In some embodiments, Figure 7 is a schematic flowchart of another thermal management method for a server provided by an embodiment of the present application. Referring to Figure 7 , the thermal management method may include the following steps.

[0132] S31. The baseboard management controller collects multi-dimensional sensor data to obtain a data packet to be processed.

[0133] Specifically, the baseboard management controller configured for a single computing node can collect the working parameters collected by sensing components such as temperature and power consumption, and perform information integration, and send them to the editable controller in the form of a "data packet to be processed, or a heat dissipation resource competition packet (which may include the temperatures and powers of the current node at multiple different times, or include the average temperature, power change rate, and priority weight of the current node)". The transmission bus between the editable controller and the baseboard management controller may not be limited to buses such as I2C or I3C.

[0134] S32. The editable controller uses a data cache module to temporarily store and transmit data.

[0135] Specifically, when the editable controller, such as an FPGA, receives the data packet to be processed transmitted by the baseboard management controller, in order to prevent the data processing speed from being less than the signal transmission speed between the editable controller and the baseboard management controller, a dynamic data cache module with a dual-port SRAM architecture is used in the editable controller to store data, avoiding the loss of subsequent transmission signals caused by too slow data processing speed, ensuring data integrity, and improving control accuracy.

[0136] S33. The editable controller uses a data preprocessing module to perform data preprocessing and feature engineering.

[0137] Specifically, after receiving all the data transmitted by the baseboard management controller, the editable controller preprocesses the data to adapt to the input of the hybrid model. Among them, in addition to aggregating the temperature data, based on the characteristics of multiple computing nodes, weights can be assigned to multiple groups of data and adaptive dynamic adjustment can be performed.

[0138] S34. The editable controller uses the current thermal management strategy prediction model module to predict the target required air volume.

[0139] Specifically, for the preprocessed data, the editable controller can perform LSTM-FNN hybrid model inference on it. According to the input relevant parameters such as temperature and power consumption, predict the possible next target air volume demand value, that is, predict the target required air volume.

[0140] S35. The editable controller performs constraint checking on the target required air volume.

[0141] Specifically, constraint checking is also called the check before output. Exemplarily, due to the limitations of the total air volume capacity of the server chassis and the fan speed under actual conditions, the generated target air volume demand value needs to undergo limit checking to obtain the optimal actual value, and the air volume is allocated to each fan through the calculated weights.

[0142] S36. The editable controller performs air volume allocation and integrates a fault tolerance mechanism and speed selection.

[0143] Specifically, the actual speed is generated based on the competition calculation of the fault tolerance mechanism and air volume allocation: a baseboard management controller status monitoring mechanism is introduced in the step of generating the speed. When the heartbeat signal of the baseboard management controller is not sent as expected, it indicates that the baseboard management controller is working abnormally. At this time, the editable controller will no longer perform air volume prediction according to the pending data packets transmitted by the baseboard management controller, but directly use a pre-built fixed or stepped heat dissipation strategy to generate a preset speed. If the status of the baseboard management controller remains normal, that is, the baseboard management controller is working properly, the speed associated with the current thermal management strategy is obtained based on the hybrid model inference.

[0144] S37. The editable controller performs speed control and receives TACH signal feedback.

[0145] Specifically, this step may include PWM control and TACH signal feedback. Exemplarily, the rotation speed selection and rotation control module of the programmable controller converts the finally calculated fan rotation speed into a PWM signal and outputs it to the fan for actual fan rotation speed control. In addition, the fan can feedback the TACH signal back to the programmable controller, and the programmable controller calculates the actual rotation speed based on this TACH signal and uses it for hybrid model update.

[0146] In an embodiment of the present application, a global distributed thermal management solution for edge intelligence based on a programmable controller, such as an FPGA, is provided. By deploying a dual-port SRAM ping-pong mechanism and an LSTM-FNN hybrid model on the FPGA, a real-time data processing capability with high throughput, low latency, and strong adaptability is achieved; based on the LSTM-FNN hybrid model, the air volume demand can be predicted in advance, improving the latency problem caused by the method of first temperature perception and then regulating the rotation speed in the related art, and reducing the risk of overheating. In addition, the multi-node air volume allocation method incorporating the Nash bidding scheme improves the PID balance problem caused by multiple fans competing and the heat source difference problem between different computing nodes. Finally, a fault tolerance and feedback mechanism is integrated to prevent the problem of untimely heat dissipation caused by the BMC of the computing node hanging; the actual rotation speed is fed back, the fan rotation speed is recorded, and the predicted expected rotation speed and the actual rotation speed of the model are displayed under the BCM network, such as a rotation speed curve, to facilitate an intuitive understanding of the current thermal management operation. Additionally, data processing based on the FPGA solves the "performance-latency-power consumption" triangular contradiction in real-time data processing, providing a highly intelligent solution for the thermal management of the next-generation intelligent system, which has both academic value (such as promoting edge AI research) and commercial potential (such as covering multi-industry application scenarios).

[0147] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented only by hardware, which is not limited herein.

[0148] An embodiment of the present application also provides a thermal management device for a server. Refer to Figure 8 , the thermal management device 40 may include: a working parameter acquisition module 41 configured to acquire the working parameters of the computing node; a data packet generation module 42 configured to generate a data packet to be processed based on the working parameters; wherein the data packet to be processed includes historical working parameters; a thermal management strategy prediction module 43 configured to predict the current thermal management strategy for at least one heat dissipation component based on the historical working parameters; wherein the current thermal management strategy is configured to control at least one heat dissipation component to dissipate heat for the current working state of the corresponding computing node.

[0149] In the embodiments of the present application, for the descriptions of the features in the corresponding embodiments of the server's thermal management device, reference can be made to the relevant descriptions in the corresponding embodiments of the server's thermal management method, which will not be elaborated herein one by one.

[0150] The embodiments of the present application further provide an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned embodiments of the server's thermal management method.

[0151] The embodiments of the present application further provide a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the steps in any of the above-mentioned embodiments of the server's thermal management method when running.

[0152] In an exemplary embodiment, the above-mentioned computer-readable storage medium may include, but is not limited to: USB flash drives, read-only memories (ROM for short), random access memories (RAM for short), mobile hard disks, magnetic disks, or optical discs and other various media that can store computer programs.

[0153] The embodiments of the present application further provide a computer program product. The above-mentioned computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps in any of the above-mentioned embodiments of the server's thermal management method.

[0154] The embodiments of the present application further provide another computer program product, including a non-volatile computer-readable storage medium. The non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps in any of the above-mentioned embodiments of the server's thermal management method.

[0155] Those skilled in the art can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0156] The above has introduced in detail the server main card, server, thermal management method, product, device and medium provided by the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A thermal management system for a server, characterized in that: The server includes at least one computing node and at least one heat dissipation component configured on the computing node; the thermal management system includes a sensing component, a baseboard management controller and an editable controller; The sensing component and the baseboard management controller are correspondingly configured at the computing node, the sensing component is connected to the baseboard management controller, and the baseboard management controller is also connected to the editable controller; The sensing component is configured to collect the operating parameters of the computing node and transmit them to the baseboard management controller; The baseboard management controller is configured to generate a data packet to be processed based on the operating parameters and transmit the data packet to the editable controller, wherein the data packet to be processed includes historical operating parameters; The editable controller is configured to predict a current thermal management strategy for the at least one heat dissipation component using a prediction model based on the historical operating parameters; the prediction model is configured to obtain output parameters based on input parameters associated with the historical operating parameters, the output parameters are associated with the current thermal management strategy, and the current thermal management strategy is configured to control at least one of the heat dissipation components to dissipate heat according to the current operating state of the corresponding computing node.

2. The thermal management system of the server according to claim 1, characterized in that: The editable controller at least includes: a data preprocessing module and a current thermal management strategy prediction model module; The current thermal management strategy prediction model module is connected to the data preprocessing module; The data preprocessing module is configured to generate the input parameters based on the historical operating parameters and transmit them to the current thermal management strategy prediction model module; The current thermal management strategy prediction model module deploys the prediction model, and the prediction model is configured to perform parameter prediction based on the input parameters to obtain the output parameters, and the output parameters are associated with the current thermal management strategy.

3. The thermal management system of the server according to claim 2, characterized in that: The heat dissipation component includes a fan; The sensing component includes at least a temperature sensing component and a power consumption sensing component, and the operating parameters of the computing node include at least temperature and power consumption; The data packet to be processed includes at least the temperature and power consumption at the current moment and at a plurality of previous historical moments; The input parameters include the current temperature gradient, the power change rate, the dynamic weight, and multiple temperature gradients and multiple power gradients within a historical preset time period; The output parameters include target air volume demand, priority weight and thermal inertia coefficient; The current thermal management strategy includes a current air volume requirement and / or a current rotation speed requirement of the at least one heat dissipation component.

4. The thermal management system of the server according to claim 3, characterized in that: The editable controller also includes: an information receiving and converting module and a dynamic data caching module; The dynamic data cache module is connected between the information transceiving and converting module and the data preprocessing module; The information receiving and converting module is configured to receive the data packet to be processed, parse the data packet to be processed to obtain target data, and transmit the target data to the dynamic data cache module; The dynamic data cache module is configured to temporarily store the target data and transmit it to the data preprocessing module based on at least a double buffer structure.

5. The thermal management system of the server according to claim 4, characterized in that: The dynamic data cache module includes an input data selection unit, a first cache unit, a second cache unit and a data output selection unit; The input end of the input data selection unit is connected to the information transceiving and converting module, the output end of the input data selection unit is connected to the input end of the first cache unit and the input end of the second cache unit, the output end of the first cache unit and the output end of the second cache unit are both connected to the data output selection unit, and the data output selection unit is also connected to the data preprocessing module; The input data selection unit is configured to direct the target data to the first cache unit and the second cache unit alternately at intervals; The first cache unit and the second cache unit are configured to alternately receive updated target data at intervals; The data output selection unit is configured to transmit the target data updated in the first cache unit and the second cache unit to the data preprocessing module alternately at intervals.

6. The thermal management system of a server according to claim 4, characterized in that: The editable controller further comprises: a speed signal reading back and feedback module, the speed signal reading back and feedback module is connected to the information receiving and transmitting and converting module; The speed signal reading and feedback module is configured to obtain the actual speed of the fan and transmit it to the information receiving and converting module; The information transceiving and converting module is further configured to forward the actual rotation speed to the baseboard management controller; The baseboard management controller is further configured with a display module, and the display module is configured to at least display the actual rotation speed.

7. The thermal management system of a server according to claim 6, characterized in that: The information receiving and converting module is further configured to forward the actual speed to the current thermal management strategy prediction model module; The current thermal management strategy prediction model module is further configured to update parameters of the configured model based on at least the actual speed.

8. The thermal management system of a server according to claim 3, characterized in that: The editable controller further comprises: a constraint checking and multi-node air volume allocation module, wherein the constraint checking and multi-node air volume allocation module is connected to the current thermal management strategy prediction model module; The constraint checking and multi-node air volume allocation module is configured to determine a competition package based on the priority weight, the thermal inertia coefficient and the current temperature gradient; and determine the current air volume requirement and / or current speed requirement of the heat dissipation components configured in different computing nodes based on the competition package, the target air volume requirement and the target constraint condition; The target constraint conditions include a total component constraint condition and a fan speed constraint condition.

9. The thermal management system of a server according to claim 8, characterized in that: The editable controller further comprises: a baseboard management controller state monitoring module, an editable controller preset speed output module and a speed selection and rotation control module, wherein the baseboard management controller state monitoring module, the editable controller preset speed output module and the constraint checking and multi-node air volume distribution module are all connected to the speed selection and rotation control module; The baseboard management controller status monitoring module is configured to obtain the working status of the baseboard management controller configured in the computing node, and the working status includes normal working and abnormal working; The editable controller preset speed output module is configured to generate a preset speed for fan control according to a preset thermal management strategy; The speed selection and rotation control module is configured to receive the working status of the baseboard management controller, and based on the working status being normal operation, use the current thermal management strategy to control the rotation of the fan; or based on the working status being abnormal operation, use the preset speed to control the rotation of the fan.

10. The thermal management system of a server according to any one of claims 3 to 9, characterized in that: The current thermal management strategy prediction model module is configured with a hybrid model of a long short-term memory network and a feedforward neural network; The LSTM network is configured to output a hidden state based on a plurality of temperature gradients and a plurality of power gradients within a historical preset time period in the input parameters; The long short-term memory network is configured to determine the target air volume requirement, the priority weight and the thermal inertia coefficient based on the hidden state and the current temperature gradient, power change rate and dynamic weight in the input parameters.

11. A server, characterized in that: A thermal management system comprising a server as described in any one of claims 1-10.

12. A thermal management method for a server, characterized in that: A thermal management system for a server according to any one of claims 1 to 10 is implemented; the thermal management method comprises: Obtaining operating parameters of the computing node; Based on the working parameters, generating a data packet to be processed; wherein the data packet to be processed includes historical working parameters; A prediction model is used to predict the current thermal management strategy for at least one heat dissipation component based on the historical operating parameters; wherein the prediction model is configured to obtain output parameters based on input parameters associated with the historical operating parameters, and the output parameters are associated with the current thermal management strategy, and the current thermal management strategy is configured to control at least one of the heat dissipation components to dissipate heat according to the current operating state of the corresponding computing node.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the thermal management method for a server as claimed in claim 12 are implemented.

14. An electronic device, characterized in that: include: Memory for storing computer programs; A processor is used to implement the steps of the thermal management method for the server as claimed in claim 12 when executing the computer program.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the thermal management method for a server as claimed in claim 12.

Images