Server heat dissipation control system, method, storage medium and program product

Abstract

The present invention discloses a server heat dissipation control system, method, storage medium and program product, which relates to the technical field of hardware heat dissipation. It includes a data acquisition layer, a resource management layer, a data reasoning layer and a decision execution layer. It can calculate the current sparsity of a neural network model through the multi-modal data and system load data of the server, then calculate the weight matrix of the neural network model through the current sparsity, use the neural network model and the weight matrix to reason about the multi-modal data, and finally generate a control signal based on the reasoning result to control the server to perform a heat dissipation action. Through the technical fusion of dynamic sparse learning and multi-modal fusion, high-performance heat dissipation control is achieved, solving the technical problem in the related art that there is a conflict between model resource occupancy and real-time performance during server heat dissipation, resulting in the inability to balance the requirements of control accuracy and real-time performance, and achieving the technical effects of optimizing sparse calculation and solving the conflict between multi-modal and embedded resources.

Description

Technical Field

[0001] The present invention relates to the technical field of hardware heat dissipation, and in particular, to a server heat dissipation control system, method, storage medium, and program product. Background Art

[0002] With the rapid development of cloud computing, AI (Artificial Intelligence), and high-performance computing, the power density of server clusters has increased significantly, and heat dissipation control has become a key challenge to ensure system stability, extend the hardware life, and reduce energy consumption.

[0003] In related technologies, a server can implement heat dissipation control based on a neural network model. However, a neural network model usually requires a large amount of memory resources. Since server heat dissipation needs to meet real-time control requirements in some scenarios and accuracy control requirements in some scenarios, there is a conflict problem between model resource occupancy and real-time performance during server heat dissipation. Summary of the Invention

[0004] The present invention provides a server heat dissipation control system, method, storage medium, and program product to at least solve the technical problems in related technologies that there is a conflict between model resource occupancy and real-time performance during server heat dissipation, resulting in the inability to balance the requirements of control accuracy and real-time performance.

[0005] The present invention provides a server heat dissipation control system, including: a data acquisition layer that acquires multimodal data and system load data of the server; a resource management layer, where the resource management layer includes a storage area in which a sparse matrix index table is pre-allocated in a memory pool, and a security area that executes a neural network model for the server's heat dissipation control task; a data inference layer that loads the neural network model from the security area, calculates the current sparsity of the neural network model according to the system load data, queries the sparse matrix index table in the storage area based on the current sparsity to obtain the current weight matrix of the neural network model, and uses the neural network model to perform inference on the multimodal data; and a decision execution layer that controls the server to perform a heat dissipation action based on a control signal generated from the inference result.

[0006] The present invention also provides a server heat dissipation control method, which performs heat dissipation control based on the above server heat dissipation control system, including: acquiring multimodal data and system load data of the server; loading the neural network model from the security area, calculating the current sparsity of the neural network model according to the system load data, querying the sparse matrix index table in the storage area based on the current sparsity to obtain the current weight matrix of the neural network model, and using the neural network model to perform inference on the multimodal data; and controlling the server to perform a heat dissipation action based on a control signal generated from the inference result.

[0007] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the above server heat dissipation control methods.

[0008] The present invention also provides a computer program product including a computer program, which, when executed by a processor, implements the steps of any one of the above server heat dissipation control methods.

[0009] Through the present invention, by using the multimodal data and system load data of the server, the current sparsity of the neural network model is calculated, then the weight matrix of the neural network model is calculated through the current sparsity, and then the multimodal data is inferred by using the neural network model and the weight matrix. Finally, based on the control signal generated by the inference result, the server is controlled to perform a heat dissipation action. Through the technical integration of dynamic sparse learning and multimodal fusion, high-performance heat dissipation control is achieved. Thus, the technical problem in the related art that there is a conflict between model resource occupancy and real-time performance during server heat dissipation, resulting in the inability to balance the requirements of control accuracy and real-time performance, can be solved, and the technical effects of optimizing heat dissipation control technology for sparse calculation, solving the conflict between multimodality and embedded resources, and balancing learning ability and security are achieved, ensuring the stability of the system, extending the hardware life, and reducing energy consumption. Description of the Drawings

[0010] To more clearly illustrate the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1 It is a schematic structural diagram of a server heat dissipation control system provided by an embodiment of the present invention;

[0012] Figure 2 It is an architecture diagram of a server heat dissipation control system provided by an embodiment of the present invention;

[0013] Figure 3 It is a flowchart of a server heat dissipation control method provided by an embodiment of the present invention;

[0014] Figure 4 It is a schematic diagram of a data acquisition process provided by an embodiment of the present invention;

[0015] Figure 5 It is a schematic diagram of a dynamic model sparse inference process provided by an embodiment of the present invention;

[0016] Figure 6Schematic diagram of the online incremental learning and security update module provided by an embodiment of the present invention;

[0017] Figure 7 Schematic flow chart of a server heat dissipation control method provided by an embodiment of the present invention.

[0018] Explanation of reference numerals: 101 - Data acquisition layer; 102 - Resource management layer; 103 - Data inference layer; 104 - Decision execution layer. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] It should be noted that in the description of the present invention, the terms "include", "comprise" or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes 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 invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0021] Currently, there are many limitations in the related technology for server heat dissipation control, which are as follows:

[0022] (1) Static policy failure: The control policy based on the temperature - fan speed mapping table (such as the IPMI (Intelligent Platform Management Interface) standard) uses fixed thresholds and cannot sense sudden changes in the load. For example, when the GPU (Graphics Processing Unit) is instantaneously fully loaded during AI training, the temperature soars at a rate of up to 10 °C / s, and the traditional solution has a risk of local overheating due to detection - response delay (usually >5 ms).

[0023] (2) PID (Proportional - Integral - Derivative) control defect: Although the classic PID algorithm can achieve closed - loop regulation, its linear characteristics are difficult to cope with the non - linear thermodynamic system of the server (such as the cubic relationship between the fan speed and the heat dissipation efficiency). Measured data shows that in the scenario of high - load mutation, the standard deviation of the temperature fluctuation of PID control ( ) can reach ±3.5℃, far exceeding industrial grade requirements ( <2℃).

[0024] (3) Energy efficiency imbalance: To avoid the risk of overheating, traditional solutions often adopt a conservative "overcooling" strategy, which causes the fan to run at high speed for a long time, and the energy consumption can account for 15%-20% of the total power consumption of the server.

[0025] In AI-driven heat dissipation control technology, the following technical problems exist:

[0026] (1) Conflict between model complexity and embedded resources: For example, the AI solution uses a 1.2M parameter LSTM (Long Short-Term Memory) network, which requires 2.3MB of storage space and 98% of the CPU (Central Processing Unit) resources, resulting in an inference delay of 62ms, far exceeding the real-time control threshold (<15ms). In addition, the large model squeezes the BMC (Baseboard Management Controller) security partition space (only 512KB is reserved), which may undermine the firmware integrity verification mechanism.

[0027] (2) Insufficient utilization of multimodal data: Existing technologies often rely on single temperature sensor data, ignoring the collaborative value of multi-dimensional information such as power consumption, airflow, and noise. For example, a local anomaly in the airflow velocity inside a server chassis may occur earlier than a temperature sensor alarm, but existing solutions lack a multimodal feature fusion mechanism, missing the opportunity for early intervention.

[0028] (3) Mismatch between model solidification and dynamic environment: The parameters of the AI model after deployment are fixed and cannot adapt to long-term environmental evolution such as hardware aging (such as dust accumulation on the heat sink leading to increased thermal resistance) and changes in the layout of the computer room (such as changes in cabinet ventilation conditions). Manual retraining of the model is required, resulting in high operation and maintenance costs.

[0029] In order to address the defects of the above-mentioned related technologies, the present invention proposes a server heat dissipation control method, electronic device, storage medium and program product, aiming to solve technical problems such as the unoptimized sparse computing of related technologies, the direct impact of multi-sensor timing asynchrony on the accuracy of control decisions, and the conflict between security and real-time performance, so as to achieve the technical effects of optimizing sparse computing, resolving the conflict between multimodality and embedded resources, and balancing learning ability and security, which will be described in detail below.

[0030] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0031] Figure 1Schematic diagram of a server heat dissipation control system provided by an embodiment of the present invention, as shown in Figure 1 shown. The system includes: a data acquisition layer 101, a resource management layer 102, a data inference layer 103, and a decision execution layer 104.

[0032] Among them, the data acquisition layer 101 is used to collect multi-modal data and system load data of the server; the resource management layer 102 includes a storage area where a sparse matrix index table is pre-allocated in the memory pool, and a security area for executing the neural network model for the server's heat dissipation control task; the data inference layer 103 is used to load the neural network model from the security area, calculate the current sparsity of the neural network model according to the system load data, query the sparse matrix index table in the storage area based on the current sparsity to obtain the current weight matrix of the neural network model, and use the neural network model to infer the multi-modal data; the decision execution layer 104 is used to control the server to execute a heat dissipation action based on the control signal generated by the inference result.

[0033] Among them, the system load data includes the computing resource occupancy rate, memory margin, etc., where the computing resource occupancy rate is the CPU occupancy rate.

[0034] It can be understood that the embodiment of the present invention includes four main levels: the data acquisition layer 101 is responsible for collecting multi-modal data and system load information of the server; the resource management layer 102 has a storage area where a sparse matrix index table is pre-allocated for storage, and also includes a security area for executing the neural network model for heat dissipation control tasks; the data inference layer 103 loads the neural network model from the security area, calculates the current sparsity according to the system load data, obtains the current weight matrix of the model by querying the sparse matrix index table, and uses this model to perform analysis and inference on the multi-modal data; finally, the decision execution layer 104 generates a control signal based on the inference result to guide the server to execute corresponding heat dissipation measures.

[0035] In the embodiment of the present invention, the multi-modal data includes sensor data of multiple components in the server. The data acquisition layer 101 includes a sensor interface and a timer. Among them, the sensor interface receives the sensor data and system load data of multiple components; the timer generates time stamps for the sensor data of multiple components and aligns the time windows of the sensor data of multiple components using the time stamps.

[0036] Among them, the sensor data of multiple components in the multi-modal data server, such as the temperature gradients, power consumption change rates, air flow speeds, and ambient noises of the CPU, GPU, memory, and hard disk, can be obtained as follows: the temperature gradient can be obtained through IPMI, the power consumption change rate can be obtained through PMBus (Power Management Bus), the air flow speed can be inversely deduced based on the current PWM (Pulse Width Modulation) duty cycle for the air flow speed inside the computer case, and the ambient noise can be obtained through an ADC (Analog-to-Digital Converter); a timer is a timing device that can provide high-precision time measurement with an accuracy of up to 1 microsecond and can generate accurate timestamps.

[0037] It can be understood that the data acquisition layer 101 in the embodiment of the present invention receives sensor data and system load data from multiple components in the server through a sensor interface, and uses a timer to add accurate timestamps at the moment when each piece of data is acquired. Then, based on these timestamps, the data of each component is aligned according to the same time window to ensure accurate synchronization even if the data sources are diverse and there may be transmission delays, providing support for subsequent comprehensive data analysis and server management.

[0038] In the embodiment of the present invention, the timer is bound to all sensor interrupt sources.

[0039] Among them, a sensor interrupt source is an interrupt request sent to the processor generated when the sensor detects that the physical quantity it monitors changes by more than a preset threshold; associating the hardware timer with all sensors allows the timer to accurately record the time point when these interrupt signals are received, thereby obtaining accurate timestamps.

[0040] In the embodiment of the present invention, the data acquisition layer 101 includes a preprocessing module and a feature extraction module. Among them, the preprocessing module processes the multi-modal data and system load data; the feature extraction module extracts features from the preprocessed multi-modal data.

[0041] Among them, the preprocessing module is used to preprocess the original multi-modal data and system load data obtained from sensors and system monitoring tools, including operations such as normalization and filtering (such as moving average denoising) to improve the data quality and prepare for subsequent analysis; the feature extraction module extracts features from the preprocessed multi-modal data, such as calculating the temperature change rate, etc.

[0042] It can be understood that in the embodiment of the present invention, in the data acquisition layer 101, a series of preprocessing operations such as normalization and filtering are performed on the multi-modal data and system load data through the preprocessing module to improve the data quality and consistency; the feature extraction module extracts features from the preprocessed multi-modal data, such as the temperature change rate, etc., ensuring the effective conversion from the original data to useful information and enhancing the intelligent level and response speed of the system.

[0043] In the embodiment of the present invention, the secure area stores the encrypted neural network model and is used for the security verification of the neural network model.

[0044] Among them, the secure area is an isolated environment that can divide the working state of the processor into secure and normal at the hardware level. In server management, the secure area can be used to protect sensitive data and code from unauthorized access; the encrypted neural network model refers to the encrypted processing of the trained neural network model to prevent unauthorized access or tampering. The encrypted model can only be loaded and used after passing the correct decryption process.

[0045] It can be understood that the embodiment of the present invention adopts a secure area, stores the encrypted neural network model therein, and is used to perform the security verification of the model. When the model needs to be used, the model needs to be verified. Once the verification passes, the model can be normally loaded and used; if the verification fails, the model loading is blocked, thereby ensuring the security and reliability of the system.

[0046] In the embodiment of the present invention, the process of security verification includes: calculating the hash value of the encrypted neural network model; performing model integrity verification based on the hash value and the pre-stored value; after the model integrity verification passes, converting the initial weight matrix of the neural network model into the compressed sparse column format and storing it in the target address of the memory pool.

[0047] Among them, the hash value is a fixed-length string generated by the hash function for the encrypted neural network model, and this string can be used as a data fingerprint to verify the integrity and consistency of the data; the pre-stored value is the pre-stored standard hash value, and the pre-stored value and the hash value are compared to achieve model integrity verification; the compressed sparse column format is CSC (Compressed Sparse Column), which is a method for storing sparse matrices, only saving non-zero elements and their corresponding row indices and column pointers to reduce memory occupancy and accelerate specific operations.

[0048] It can be understood that in the security verification process of the embodiments of the present invention, first, the hash value of the encrypted neural network model is calculated, and the integrity of the model is verified by comparing it with the pre-stored correct hash value. If the two are the same, it indicates that the verification is passed. After the verification is passed, the initial weight matrix of the neural network model is converted into a compressed sparse column format to optimize the storage and calculation efficiency. Subsequently, the converted sparse matrix is stored in the specified target address in the memory pool to provide support for subsequent steps. Through the above operations, it is ensured that the model not only maintains its original state without being tampered with, but also can efficiently utilize system resources, enhancing the security and performance of the overall system.

[0049] In the embodiments of the present invention, the data inference layer 103 includes an adjustment module. Among them, the adjustment module loads the encrypted neural network model, receives the system load data of the data acquisition layer 101, and calculates the current sparsity of the neural network model according to the system load data.

[0050] Among them, the sparsity refers to the proportion of non-zero parameters in the neural network model. A higher sparsity means that more parameters are set to zero, thereby reducing the amount of calculation and storage requirements. The sparsity is usually expressed as a value between 0 and 1, where 0 represents completely dense and 1 represents completely sparse; the current sparsity can be calculated according to the calculation resource occupancy rate using the following formula: , is the current sparsity, is the sigmoid function, which is used to map the x value to the interval (0, 1) to achieve a smooth transition. .

[0051] It can be understood that in the embodiments of the present invention, the adjustment module first loads the decrypted secure neural network model and receives the system load data from the data acquisition layer 101. Based on these real-time system load information, the adjustment module dynamically calculates the optimal sparsity value suitable for the current workload. It can be seen from the above sparsity calculation formula that the sparsity of the neural network model is in a direct proportional relationship with the calculation resource occupancy rate, that is, when the calculation resource occupancy rate is higher, the sparsity of the neural network model is also higher. Conversely, when the calculation resource occupancy rate is lower, the sparsity of the neural network model is also lower. For example, when the calculation resource occupancy rate is relatively high, the server is under high load, and at this time, the calculated sparsity is 90%, sacrificing part of the accuracy to ensure real-time performance, thereby achieving a trade-off between accuracy and latency. When the calculation resource occupancy rate is relatively low, the server is under low load, and at this time, the calculated sparsity is 30%, appropriately increasing the amount of calculation and storage requirements, thereby improving the model accuracy and achieving the optimization of resource utilization.

[0052] In an embodiment of the present invention, the data inference layer 103 includes a calculation module. The calculation module converts the current weight matrix into a compressed sparse column format and uses the target instruction set to implement parallel multiply-accumulate operations on the current weight matrix.

[0053] Among them, the target instruction set refers to the instruction set supported by a specific hardware platform, such as the NEON SIMD (Single Instruction Multiple Data Streams) extension in the ARM architecture. These instruction sets can accelerate specific types of computational operations, such as matrix multiplication.

[0054] It can be understood that in the data inference layer 103 of the embodiment of the present invention, the calculation module first converts the current weight matrix of the neural network model into a compressed sparse column format to optimize storage and computational efficiency; then, using the instruction set provided by the target hardware platform, parallel multiply-accumulate operations on the converted weight matrix are implemented, which improves the speed of matrix operations and also reduces energy consumption by efficiently utilizing hardware resources, ensuring high-performance operation of the system.

[0055] In an embodiment of the present invention, the data inference layer 103 includes a fusion module. The fusion module fuses multimodal data into a multi-dimensional feature vector, inputs the multi-dimensional feature vector into the neural network model, and the neural network model outputs an inference result.

[0056] Among them, fusing multimodal data into a multi-dimensional feature vector, for example, splicing the temperature gradient, power consumption change rate, and air flow velocity in the multimodal data into a three-dimensional feature vector.

[0057] It can be understood that in the data inference layer 103 of the embodiment of the present invention, the fusion module is used to integrate multimodal data into a multi-dimensional feature vector. Then, this multi-dimensional feature vector is fed into the neural network model, and the neural network model deeply analyzes it and generates corresponding inference results, enabling the system to optimize the heat dissipation control strategy based on accurate predictions and ensuring efficient and stable operation.

[0058] In an embodiment of the present invention, the decision execution layer 104 includes an optimization control module, a hardware driver module, and a feedback adjustment module. The optimization control module takes multiple factors such as temperature stability, fan life, and system energy efficiency as optimization goals, generates a first drive signal according to the control signal and the error signal fed back by the feedback adjustment module, and sends the first drive signal to the hardware driver module; the hardware driver module drives the heat dissipation array to perform heat dissipation actions based on the first drive signal.

[0059] Among them, the heat dissipation array refers to a set of fans or other cooling devices used for heat dissipation inside the server.

[0060] It can be understood that in the decision execution layer 104 of the embodiments of the present invention, the optimization control module generates a first drive signal based on multiple optimization objectives such as temperature stability, fan life, and system energy efficiency, in combination with the inference result and the error signal provided by the feedback adjustment module. The hardware drive module executes specific heat dissipation measures by directly controlling the actions of the heat dissipation array according to this drive signal, ensuring that the system can dynamically respond to real-time environmental changes, while taking into account efficient heat dissipation, extending the service life of the fan, and reducing energy consumption, achieving intelligent and efficient server heat dissipation management.

[0061] In the embodiments of the present invention, the decision execution layer 104 further includes an abnormal fusing mechanism. Among them, the abnormal fusing mechanism includes: obtaining the current temperature of the server; if the current temperature exceeds the temperature threshold, generating a second drive signal; and sending the second drive signal to the hardware drive module, and the hardware drive module drives the heat dissipation array to execute a heat dissipation action based on the second drive signal.

[0062] Among them, the temperature threshold is a preset upper limit of the safe temperature. When the actual temperature of the server exceeds the temperature threshold, corresponding heat dissipation actions will be triggered. The temperature threshold is specifically set according to the actual situation and will not be specifically limited here.

[0063] It can be understood that in the decision execution layer 104 of the embodiments of the present invention, the current temperature of the server is monitored in real time through the abnormal fusing mechanism, that is, once it is detected that the temperature exceeds the preset temperature threshold, a second drive signal will be immediately generated, and the second drive signal will be sent to the hardware drive module, prompting the heat dissipation array to quickly execute an enhanced heat dissipation action, realizing a rapid response when the server temperature rises abnormally, executing an enhanced heat dissipation action, effectively preventing the hardware from being damaged due to overheating, and ensuring the stability and security of the system.

[0064] In the embodiments of the present invention, the resource management layer 102 responds to the server heat dissipation control requirement, starts the heat dissipation control thread of the server, and the heat dissipation control thread is used to execute the heat dissipation control task of the server and set the priority of the heat dissipation control task.

[0065] Among them, the heat dissipation control thread is used to monitor and adjust the control tasks of the server heat dissipation system; set the priority of the heat dissipation control thread. For example, set the priority of the heat dissipation control thread to the highest real-time priority, so that it can run preferentially, ensuring that once an abnormal temperature or other heat dissipation requirements are detected, actions can be taken immediately.

[0066] It can be understood that, in order to ensure that the heat dissipation operation can respond quickly when necessary and avoid hardware damage or other performance problems caused by overheating, the resource management layer 102 will start a dedicated heat dissipation control thread. This thread is responsible for executing specific heat dissipation tasks, such as monitoring the temperature and adjusting the fan speed. At the same time, the resource management layer 102 will set the priority of this thread to ensure that the heat dissipation control task can still be processed in a timely manner under high load, ensuring that the server can maintain good heat dissipation under various working conditions, avoiding the risks brought by overheating, and ensuring the stability and reliability of the system. By reasonably allocating resources and task priorities, the system can efficiently respond to real-time heat dissipation requirements and maintain the best operating state.

[0067] In the embodiment of the present invention, before starting the heat dissipation control thread of the server, the resource management layer 102 sets an interrupt response threshold for triggering the heat dissipation control thread and binds the interrupt to the target processing core of the server.

[0068] Among them, the interrupt response threshold needs to match the feedback requirements of the fan Tachometer signal (the Tachometer signal is a pulse signal output by a sensor built in the fan and is proportional to the rotation speed of the fan), and is specifically set according to actual requirements, which will not be specifically limited here. Binding the interrupt to the target processing core of the server can reduce the cross-core switching delay.

[0069] It can be understood that, in order to match the feedback requirements of the fan Tachometer signal, before starting the heat dissipation control thread of the server, it is necessary to identify the characteristics of the Tachometer signal provided by the fan, and accordingly set the interrupt response threshold, bind the interrupt to the target processing core of the server, and reduce the cross-core switching delay.

[0070] In the embodiment of the present invention, before starting the heat dissipation control thread of the server, the resource management layer 102 configures the direct memory access channel and uses the direct memory access channel to achieve the transmission of sensor data to the hardware acceleration engine corresponding to the sensor interface.

[0071] Among them, the direct memory access channel hardware acceleration engine refers to the part used to accelerate the analysis and processing of sensor data.

[0072] It can be understood that, before starting the heat dissipation control thread of the server, the resource management layer 102 in the embodiment of the present invention configures the direct memory access channel and uses the direct memory access channel to directly transmit the sensor data to the hardware acceleration engine corresponding to the sensor interface, avoiding the CPU as an intermediary, enabling the sensor data to be efficiently transmitted and quickly processed by the hardware acceleration engine, thereby improving the overall data processing efficiency, reducing the CPU burden, and ensuring that the heat dissipation control task can respond in a timely manner, guaranteeing the stability and efficient operation of the system.

[0073] In an embodiment of the present invention, the server heat dissipation control system further includes: a model update layer, where the model update layer includes a monitoring module, an update module, and a secure hot-swap module. The monitoring module obtains monitoring data during the heat dissipation control process of the server and predicts the performance degradation rate of the neural network model based on the monitoring data; the update module performs gradient update on the parameters of the neural network model when the performance degradation rate of the neural network model is greater than the update threshold; the secure hot-swap module realizes the switch between the neural network model before update and the neural network model after update based on integrity verification; the update module clips the updated gradient during the update to limit the update amplitude.

[0074] Among them, the monitoring data refers to the data collected from each component of the server during the heat dissipation control process, including temperature, power consumption, air flow speed, fan speed, etc., and is used to evaluate the current heat dissipation state and system performance; the performance degradation rate: measures the degree of decline in the prediction accuracy and effect of the model over time. It is usually calculated by monitoring the difference between the actual operation data and the model prediction results; gradient clipping can avoid the problem of gradient explosion by restricting the maximum value of the gradient, thereby maintaining the stability of model training; the update threshold is specifically set according to actual requirements and is not specifically limited here.

[0075] It can be understood that in the server heat dissipation control system of the embodiment of the present invention, the model update layer first collects the monitoring data during the heat dissipation process by the monitoring module and predicts the performance degradation rate of the neural network model based on this. When the performance degradation rate exceeds the set update threshold, the update module starts the parameter update process. During this period, the gradient clipping technology is used to limit the update amplitude. At the same time, the secure hot-swap module ensures the correctness and security of the new model based on integrity verification and realizes the seamless switch between the old and new models after passing the verification, ensuring that the model can be dynamically adjusted and optimized according to the actual situation without affecting the normal operation of the system, and maintaining efficient and stable heat dissipation control performance.

[0076] The server heat dissipation control system proposed according to the embodiments of the present invention can, through the data acquisition layer, resource management layer, data inference layer, and decision execution layer, utilize the multi-modal data and system load data of the server to calculate the current sparsity of the neural network model, then calculate the weight matrix of the neural network model through the current sparsity, then use the neural network model and the weight matrix to infer the multi-modal data, and finally, based on the control signal generated by the inference result, control the server to perform heat dissipation actions. Through the technical integration of dynamic sparse learning and multi-modal fusion, a lightweight time series alignment algorithm is designed, a secure learning container is constructed, high-performance heat dissipation control is achieved, and the technical effects of optimizing sparse calculation in heat dissipation control technology, solving the conflict between multi-modal and embedded resources, and balancing learning ability and security are achieved, ensuring the stability of the system, extending the hardware life, and reducing energy consumption.

[0077] The server heat dissipation control system will be further described below through a specific embodiment.

[0078] The architecture of the server heat dissipation control system in this embodiment is as Figure 2 shown, mainly consisting of a multi-modal data acquisition layer, a dynamic learning sparse engine, a real-time decision execution layer, an online incremental learning module, and a hardware resource management layer, specifically as follows:

[0079] I. Multi-modal data acquisition layer

[0080] Heterogeneous sensor data acquisition: Real-time acquisition of data such as the temperature (obtained through IPMI), power consumption (obtained through PMBus), chassis air flow velocity (inferred and calculated through PWM), and ambient noise (obtained through ADC) of each component (CPU, GPU, memory, hard disk) of the server.

[0081] Hardware-level time alignment: Use the built-in hardware timer (with an accuracy of 1 μs) to timestamp all sensor data to solve the asynchronous sampling problem (error < 1 ms).

[0082] Data preprocessing: Normalize, filter (such as moving average denoising), and extract features (such as calculating the temperature change rate dT / dt) from the original data.

[0083] II. Dynamic sparse learning engine

[0084] Variable sparsity inference: Dynamically adjust the neural network sparsity (30% - 90%) according to the real-time system load (CPU occupancy rate, memory remaining), balancing calculation accuracy and latency.

[0085] Sparse matrix accelerated calculation: Convert the model weights into compressed sparse column format and use instruction sets (such as AST2600 NEON SIMD instruction set, etc.) to implement parallel multiply-accumulate operations, improving the sparse calculation efficiency by 3.2 times.

[0086] Multi-modal feature fusion: The temperature gradient, power consumption change rate, and air flow velocity are concatenated into a three-dimensional feature vector and input into a dynamic sparse network for joint inference.

[0087] III. Real-time decision execution layer

[0088] Multi-objective optimization control: Based on the inference results, a PWM control signal is generated, and at the same time, the temperature stability (σ < 2°C), fan life (rotation speed change rate < 200 rpm / s), and system energy efficiency (PUE < 1.1) are optimized.

[0089] Hardware PWM direct drive: The signal is directly output to the fan array through a hardware PWM controller, and the end-to-end delay is < 1 ms.

[0090] Abnormal fuse mechanism: When the detected temperature exceeds the limit (e.g., > 95°C), the full-speed heat dissipation mode is directly triggered bypassing the algorithm.

[0091] IV. Online incremental learning module

[0092] Sliding window trigger: Continuously monitor the control effect (such as temperature fluctuation, energy consumption), and activate learning when the performance decay within the window (30 seconds) exceeds the threshold.

[0093] Selective parameter update: Only update the gradients of the key neuron weights (accounting for < 5%), reducing the computational amount by 90%.

[0094] Safe model hot swapping: Use a double-buffer memory design to achieve seamless switching between the old and new models, and combine TrustZone (secure area) to encrypt and store model parameters.

[0095] V. Hardware resource management layer

[0096] Memory pool pre-allocation: Pre-allocate a sparse matrix index table and a non-zero weight storage area in SRAM (Static Random Access Memory), avoiding DDR4 (Double Data Rate 4) access latency.

[0097] Task priority scheduling: The heat dissipation control task is set to the highest real-time priority (priority = 0), and can preempt other background services of BMC (such as logging).

[0098] Figure 3 For the flowchart of the server heat dissipation control method provided in this embodiment, as Figure 3 shown, the method includes the following steps:

[0099] Step 1: System initialization phase:

[0100] 1. Hardware resource configuration

[0101] SRAM pre - allocation: Dedicated SRAM area: 512KB, storing sparse index (120KB, CSC format) and weights (360KB, FP16 (Half Precision Floating Point) quantization), reserving 32KB redundancy. The dynamic memory occupancy during model operation does not exceed the reserved redundancy space (32KB).

[0102] DDR4 double - buffer pool: Pre - allocate a 1MB area for redundant backup during model hot - switching.

[0103] NEON instruction configuration: Enable the NEON SIMD (NEON refers to an Advanced SIMD (Single Instruction Multiple Data) technology in the ARM (Advanced RISC Machine architecture) architecture, which is an SIMD architecture extension for accelerating multimedia and signal processing applications; SIMD is Single Instruction Multiple Data, a parallel computing technology that allows one instruction to process multiple data points simultaneously, thus improving the efficiency of specific types of computing tasks) extension of the processor core, set the floating - point mode to half - precision (FP16) to adapt to the quantized weight format of the dynamic sparse model.

[0104] Configure the DMA (Direct Memory Access) channel to achieve zero - copy transfer of sensor data to the hardware acceleration engine.

[0105] Enable the TrustZone security region. The TrustZone security region aims to provide a hardware - level security environment for the device. By creating two independent execution regions at the processor level, namely the secure region and the non - secure region, it realizes the protection of sensitive data and code, ensuring strict isolation between the secure region and the non - secure region, and dividing the secure area (storing encrypted models) and the non - secure area (processing regular tasks).

[0106] Peripheral interface initialization: Configure the IPMI interface, enable the PMBus interface, initialize the PWM controller, and start the ADC channel.

[0107] Timer synchronization: Configure the TIM3 (Timer 3) hardware timer and bind it to all sensor interrupt sources.

[0108] This hardware resource configuration is highly compatible and embedded-friendly, with a model memory footprint of less than 512KB, adapting to DDR4 memory limitations. SRAM pre-allocation mitigates memory fragmentation, allowing for 72 hours of continuous operation without memory leaks. A security enhancement mechanism, based on TrustZone, isolates model encrypted storage from the incremental learning process, preventing malicious code injection attacks. A CRC32 hardware engine verifies model integrity, resulting in a low bit error rate.

[0109] 2. Model loading and verification

[0110] Load a pre-trained dynamic sparse model (initial sparsity 70%, number of parameters <100K).

[0111] Use the CRC32 (Cyclic Redundancy Check 32) hardware engine.

[0112] Calculate the model hash value and compare it with the pre-stored value to complete the integrity check.

[0113] Convert the model weights to CSC format and store them in the specified SRAM address (0x1E720000).

[0114] Through the coordinated optimization of a dynamic sparse learning engine and hardware acceleration, end-to-end control latency is achieved to <15ms (when sparsity >60%), a significant improvement over traditional solutions. This effectively handles sudden GPU / CPU load fluctuations (e.g., a power surge of 200W / s). Fast multimodal data fusion: A hardware timer is used to time-align multi-sensor data (with an error of <1ms), eliminating the 10ms error caused by traditional software interpolation and increasing the lead time for anomaly detection to over 200ms.

[0115] Step 2: Real-time task scheduling and interrupt optimization

[0116] 1. Real-time kernel configuration

[0117] Apply the PREEMPT_RT (Real-Time Preemption Patch) patch to the Linux kernel to enable fully preemptive scheduling.

[0118] Set the priority of the heat dissipation control thread to ensure the highest real-time performance:

[0119] 2. Interrupt Binding and Optimization

[0120] Bind the TIM3 interrupt (IRQ (Interrupt Request) 45) to CPU Core 0 to reduce cross-core switching latency:

[0121] Set the PWM controller interrupt response threshold to 10 μs to match the feedback requirements of the fan Tachometer signal.

[0122] 3. Main control thread starts

[0123] Create a real-time control thread and loop to wait for the TIM3 interrupt to trigger (1 ms period).

[0124] Step 3: Multimodal data acquisition and synchronization

[0125] The data acquisition process is as Figure 4 shown below:

[0126] 1. Sensor data reading

[0127] Temperature: Read the CPU temperature value through IPMI instructions.

[0128] Power consumption: Read the instantaneous power through PMBus instructions.

[0129] Airflow velocity: Invert the CMF (Cubic Feet per Minute) value based on the current PWM duty cycle: , where is the PWM duty cycle, that is, the proportion of the high-level time in the PWM signal to the entire cycle.

[0130] Noise analysis: After ADC sampling, perform a 256-point FFT (Fast Fourier Transform) using a Hamming window to extract the energy in the 50 Hz - 1 kHz frequency band.

[0131] 2. Hardware timestamp synchronization

[0132] Read the TIM3 counter value (1 μs accuracy) each time the data arrives and store it in a timestamped structure.

[0133] Time window alignment: Use a 1 ms window to discard data packets with a time deviation exceeding ±0.5 ms.

[0134] 3. Data buffering and filtering

[0135] Double buffer queue: Use a lock-free circular buffer to store the aligned data.

[0136] Moving average filtering: Apply a 5-point window filter to the temperature data and exponential smoothing (α = 0.2) to the power consumption data.

[0137] Fuse the multimodal data into a multi-dimensional feature vector. For example, concatenate the temperature gradient, power consumption change rate, and airflow velocity in the multimodal data into a three-dimensional feature vector.

[0138] Step 4: Dynamic Sparse Model Inference

[0139] The specific inference flowchart is as Figure 5 shown, and specifically includes:

[0140] 1. Dynamic Sparsity Adjustment:

[0141] Read the CPU utilization rate in real time through the PMU counter, calculate the sparsity. In the formula, Ucpu is the real-time CPU utilization percentage (0 - 100%), and it needs to be normalized before input; the sigmoid function is used to map the x value to the interval (0, 1) to achieve a smooth transition. .

[0142] 2. Sparse Matrix Loading: Load the pre-trained CSC format model according to the current sparsity:

[0143] Column Pointer Array (col_ptr): 32-bit integer, recording the starting position of non-zero elements in each column.

[0144] Row Index Array (row_ind): 16-bit integer, storing the row numbers in a compressed manner.

[0145] Non-zero Weights (values): FP16 format, loaded into SRAM aligned by 4KB.

[0146] 3. NEON Instruction Accelerated Computation

[0147] Use ARM assembly to optimize the sparse matrix multiplication, such as using vld1.16 to load weights, Vmla.f16 for multiply-accumulate operations, and loop unrolling x4 optimization to achieve NEON instruction accelerated computation.

[0148] Step 5: Control Signal Generation and Closed-loop Feedback

[0149] 1. Multi-objective Optimization Solution

[0150] Cost Function Minimization:

[0151] Among them, σT is the standard deviation of temperature, reflecting the temperature fluctuation. Reducing σT can improve the system stability; Δrpm is the rotational speed change amount, indicating, for example, the fan speed adjustment amplitude. Reducing Δrpm can extend the device life; PUE (Power Usage Effectiveness) is the energy efficiency index of the data center, which is the ratio of the total energy consumption to the IT equipment energy consumption. The closer PUE is to 1, the higher the energy efficiency.

[0152] Gradient Descent Iteration: Iterate 3 times with a fixed step size (0.01) to solve the optimal PWM duty cycle.

[0153] 2. PWM Signal Output

[0154] Write the calculated duty cycle into the PWM controller register.

[0155] 3. Closed-loop Speed Regulation

[0156] Read the fan Tachometer signal and calculate the speed error: , where is the target speed, i.e., the desired speed set by the system according to requirements; is the actual speed, i.e., the current speed of the fan measured in real time through the Tachometer signal; is the speed error, which is the difference between the target speed and the actual speed, reflecting the speed deviation that the system needs to compensate.

[0157] PID Regulation.

[0158] Through the PID control algorithm, adjust the control output by combining the speed error to achieve the desired speed, realize closed-loop feedback regulation, and ensure that the system can dynamically respond and stabilize at the target state.

[0159] Through the multi-objective optimization strategy (joint optimization of temperature stability, fan life, and system energy efficiency), under the same heat dissipation requirements, the fan energy consumption ratio is reduced compared with the traditional solution. The energy efficiency ratio of sparse computing is improved: the sparse matrix operation accelerated by NEON instructions enables the computing power per unit power consumption to reach 3.2 times that of traditional dense computing, which is suitable for low-power embedded scenarios.

[0160] Step Six: Online Incremental Learning and Secure Update

[0161] The process of online incremental learning and secure update is as Figure 6 shown, specifically including:

[0162] 1. Performance Monitoring Trigger

[0163] Calculate the temperature prediction RMSE (Root Mean Square Error) every 30 seconds:

[0164]

[0165] where N = 3000, which is the total number of time steps. For example, sampling once every 30 seconds, a total of 3000 data points; is the actual temperature value at time t; is the predicted temperature value at time t.

[0166] If the RMSE exceeds the baseline by 5%, trigger incremental learning.

[0167] 2. Lightweight Parameter Update

[0168] Gradient mask generation: Select the top 5% of the parameters with the absolute value of the gradient after backpropagation (sorted by L1 norm).

[0169] Gradient clipping: Limit the update amplitude (Δw ∈ [-0.1, 0.1]) to prevent overfitting.

[0170] Generate an incremental parameter package and encrypt the parameter package.

[0171] 3. Safe hot swapping

[0172] Double-buffer switching: Atomically switch the model pointer (the old model is retained for 3 cycles for rollback).

[0173] Integrity check: Calculate the model hash (polynomial 0x04C11DB7), which takes 0.8 ms. If the check fails, roll back to the old model, issue an alarm, and record a SEL (System Event Log) event; if the check passes, switch the model pointer and update the model.

[0174] Step 7: Exception handling and fusing

[0175] 1. Temperature overlimit fusing

[0176] Hardware comparator: Monitor the temperature signal in real time. Trigger conditions:

[0177] 85°C: Record an IPMI SEL event and trigger an alarm.

[0178] 95°C: Force PWM = 100% (full-speed heat dissipation).

[0179] 105°C: Pull down the GPIO (General-Purpose Input / Output) pin 12, and cut off the power supply through the CPLD (Complex Programmable Logic Device) (delay < 3 ms).

[0180] 2. Safe recovery mechanism

[0181] After fusing, the temperature needs to drop below 75°C (hysteresis of 80°C - 5°C) before switching back to the normal model.

[0182] After the fusing signal is triggered, keep the low level for at least 10 ms to avoid accidental power-off caused by transient temperature fluctuations.

[0183] Through online incremental learning, it is possible to support online updates of model parameters (only the top 5% of critical weights are updated), adapt to long-term changes such as hardware aging and data center layout changes without downtime, and the performance degradation rate after model update is <1%. Dynamic sparsity adjustment: adjusts the model sparsity in real time according to the system load (30% - 90%). When the load is high, it gives priority to ensuring latency (sparsity 90%); when the load is low, it improves accuracy (sparsity 30%), achieving optimal allocation of computing resources and enabling the model to have long-term environmental adaptability.

[0184] Figure 7 FIG. 4 is a schematic flowchart of a server heat dissipation control method provided by an embodiment of the present invention. The method performs heat dissipation control based on the above-mentioned server heat dissipation control system, as Figure 7 shown. The method specifically includes the following steps:

[0185] In step S201, multi-modal data and system load data of the server are acquired.

[0186] It can be understood that in the embodiment of the present invention, first, multi-modal data composed of temperature gradients, power consumption change rates, air flow velocities, and environmental noises of various components of the server such as the CPU, GPU, memory, and hard disk, as well as system load data such as CPU occupancy and memory remaining are acquired through various sensors or interface units, etc.

[0187] In step S202, the current sparsity of the neural network model is calculated according to the system load data.

[0188] Among them, the sparsity refers to the proportion of non-zero parameters in the neural network model. A higher sparsity means that more parameters are set to zero, thus reducing the computational amount and storage requirements; calculating the current sparsity of the neural network model according to the system load data includes: identifying the computing resource occupancy rate in the system load data; calculating the current sparsity according to the computing resource occupancy rate.

[0189] It can be understood that the current sparsity in the embodiment of the present invention is obtained by calculating the computing resource occupancy rate. The computing resource occupancy rate percentage is normalized and input into the above formula to calculate the current sparsity. The sparsity of the neural network model is in a direct proportional relationship with the computing resource occupancy rate, that is, when the computing resource occupancy rate is higher, the sparsity of the neural network model is also higher; conversely, when the computing resource occupancy rate is lower, the sparsity of the neural network model is also lower. For example, when the computing resource occupancy rate is high, the server is in a high load state, and at this time, the calculated sparsity is 90%, sacrificing some accuracy to ensure real-time performance, thereby achieving a trade-off between accuracy and latency. When the computing resource occupancy rate is low, the server is in a low load state, and at this time, the calculated sparsity is 30%, appropriately increasing the computational amount and storage requirements, thereby improving the model accuracy.

[0190] In step S203, based on the current sparsity, the weight matrix of the neural network model is calculated, and the neural network model and the weight matrix are used to perform inference on the multimodal data.

[0191] Among them, the weight matrix is a set of connection weights between neural network layers, representing the connection strength between each neuron; the neural network model is sparsified using the current sparsity. During the sparsification process, many elements in the weight matrix are set to zero, forming a sparse matrix; using the neural network model and the weight matrix to perform inference on the multimodal data includes: extracting multimodal features in the multimodal data and fusing the multimodal features; inputting the fused multimodal features into the neural network model, and the neural network model outputs the inference result, and the weight matrix is used to accelerate the inference process of the neural network model.

[0192] It can be understood that in the embodiment of the present invention, the neural network model and its corresponding sparse weight matrix are used to perform inference on the multimodal data. First, the obtained multimodal data needs to be extracted, the extracted various types of features are fused, and the fused multimodal feature vector is input into the above-mentioned neural network model with dynamically adjusted current sparsity. The inference result is output through the neural network model, reducing the computational complexity while ensuring the accuracy and real-time performance of the inference.

[0193] In step S204, based on the control signal generated from the inference result, the server is controlled to perform a heat dissipation action.

[0194] It can be understood that in the embodiment of the present invention, the current heat state result of the server is obtained through inference, and specific control signals are generated according to the inference result, such as adjusting the PWM duty cycle of the fan to increase or decrease the fan speed, etc. The server performs corresponding heat dissipation actions according to the received control signals, controls the speed of the fan, thereby effectively managing the internal temperature and ensuring that the hardware operates under the best working conditions.

[0195] For the description of the features in the corresponding embodiment of the server heat dissipation control method, reference can be made to the relevant description in the corresponding embodiment of the server heat dissipation control system, which will not be elaborated here one by one.

[0196] The server heat dissipation control method proposed according to the embodiments of the present invention utilizes the multi-modal data and system load data of the server to calculate the current sparsity of the neural network model, then calculates the weight matrix of the neural network model through the current sparsity, then uses the neural network model and the weight matrix to perform inference on the multi-modal data, and finally controls the server to perform a heat dissipation action based on the control signal generated by the inference result. Through the integration of dynamic sparse learning and multi-modal fusion technologies, a lightweight time series alignment algorithm is designed, a secure learning container is constructed, high-performance heat dissipation control is achieved, and the technical effects of optimizing sparse computing in heat dissipation control technology, solving the conflict between multi-modal and embedded resources, and balancing learning ability and security are achieved, ensuring the stability of the system, extending the hardware life, and reducing energy consumption.

[0197] 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 by hardware, but in many cases, the former is a better implementation method.

[0198] The embodiments of the present invention also provide a computer-readable storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above embodiments of the server heat dissipation control method when running.

[0199] In an exemplary embodiment, the above computer-readable storage medium may include, but is not limited to: USB flash drive, read-only memory (ROM for short), random access memory (RAM for short), mobile hard disk, magnetic disk or optical disc, etc., various media that can store computer programs.

[0200] The embodiments of the present invention also provide a computer program product, the above computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps in any one of the above embodiments of the server heat dissipation control method.

[0201] The embodiments of the present invention also 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 one of the above embodiments of the server heat dissipation control method.

[0202] Those skilled in the art may 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 both. 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 invention.

[0203] The above has introduced in detail a server heat dissipation control method, an electronic device, a storage medium, and a program product provided by the present invention. Specific examples are used herein to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A server heat dissipation control system, characterized in that, Including: A data acquisition layer that acquires multi-modal data and system load data of the server; A resource management layer, where the resource management layer includes a storage area with a sparse matrix index table pre-allocated in a memory pool and a security area of a neural network model that performs the heat dissipation control task of the server; A data inference layer that loads the neural network model from the security area, calculates the current sparsity of the neural network model according to the system load data, queries the sparse matrix index table in the storage area based on the current sparsity to obtain the current weight matrix of the neural network model, and uses the neural network model to perform inference on the multi-modal data. Among them, the data inference layer includes an adjustment module that loads an encrypted neural network model, receives the system load data of the data acquisition layer, and calculates the current sparsity of the neural network model according to the system load data; A decision execution layer that controls the server to perform a heat dissipation action based on a control signal generated from an inference result.

2. The server heat dissipation control system according to claim 1, wherein The multi-modal data includes sensor data of multiple components in the server. The data acquisition layer includes a sensor interface and a timer, where The sensor interface receives the sensor data of the multiple components and the system load data; The timer generates a timestamp for the sensor data of the multiple components and aligns the time windows of the sensor data of the multiple components using the timestamp.

3. The server heat dissipation control system according to claim 2, wherein The timer is bound to all sensor interrupt sources.

4. The server heat dissipation control system according to claim 1 or 2, characterized in that The data acquisition layer includes a preprocessing module and a feature extraction module, where The preprocessing module preprocesses the multi-modal data and the system load data; The feature extraction module extracts features from the preprocessed multi-modal data.

5. The server heat dissipation control system according to claim 1, wherein The security area stores an encrypted neural network model and is used for the security verification of the neural network model.

6. The server heat dissipation control system according to claim 5, characterized in that, The process of the security verification includes: Calculating the hash value of the encrypted neural network model; Performing model integrity verification based on the hash value and a pre-stored value; After the model integrity verification passes, converting the initial weight matrix of the neural network model into a compressed sparse column format and storing it at the target address in the memory pool.

7. The server heat dissipation control system according to claim 6, characterized in that The data inference layer includes a calculation module, where the calculation module converts the current weight matrix into a compressed sparse column format and realizes the parallel multiply-accumulate operation of the current weight matrix using a target instruction set.

8. The server heat dissipation control system according to claim 7, wherein, The data inference layer includes a fusion module, where the fusion module fuses the multi-modal data into a multi-dimensional feature vector, inputs the multi-dimensional feature vector into the neural network model, and the neural network model outputs an inference result.

9. The server heat dissipation control system according to claim 1, characterized in that The decision execution layer includes an optimization control module, a hardware driver module, and a feedback adjustment module, where The optimization control module takes multiple of temperature stability, fan life, and system energy efficiency as optimization goals, generates a first drive signal according to the control signal and an error signal fed back by the feedback adjustment module, and sends the first drive signal to the hardware driver module; The hardware driver module drives a heat dissipation array to perform a heat dissipation action based on the first drive signal.

10. The server heat dissipation control system according to claim 9, characterized in that, The decision execution layer further includes an abnormal fusing mechanism, where the abnormal fusing mechanism includes: Obtain the current temperature of the server; If the current temperature exceeds the temperature threshold, generate a second driving signal; Send the second driving signal to the hardware driving module, and the hardware driving module drives the heat dissipation array to perform a heat dissipation action based on the second driving signal.

11. The server heat dissipation control system according to claim 1, wherein The resource management layer responds to the server heat dissipation control requirement, starts the heat dissipation control thread of the server, the heat dissipation control thread is used to execute the heat dissipation control task of the server, and sets the priority of the heat dissipation control task.

12. A server heat dissipation control method, characterized in that, The method performs heat dissipation control based on the server heat dissipation control system according to any one of claims 1-11, where the method includes: Obtain the multimodal data and system load data of the server; Load the neural network model from the security area, calculate the current sparsity of the neural network model according to the system load data, query the sparse matrix index table in the storage area based on the current sparsity, obtain the current weight matrix of the neural network model, and use the neural network model to perform inference on the multimodal data; Based on the control signal generated by the inference result, control the server to perform a heat dissipation action.

13. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, where the computer program, when executed by a processor, implements the steps of the server heat dissipation control method according to claim 12.

14. A computer program product, comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the server heat dissipation control method according to claim 12.

Images