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

Abstract

The invention discloses a server heat dissipation control system and method, a storage medium and a program product, and relates to the technical field of hardware heat dissipation, the server heat dissipation control system comprises a data acquisition layer, a resource management layer, a data reasoning layer and a decision execution layer, the current sparseness of a neural network model can be calculated through multi-modal data and system load data of a server, and the current sparseness of the neural network model is calculated; the method comprises the following steps: firstly, calculating the current sparseness of a neural network model, then calculating a weight matrix of the neural network model through the current sparseness, reasoning the multi-modal data by using the neural network model and the weight matrix, finally, generating a control signal based on a reasoning result, controlling a server to execute a heat dissipation action, and realizing high-performance heat dissipation control through dynamic sparse learning multi-modal fusion technology fusion. The technical problem that the requirements of control precision and real-time performance cannot be considered at the same time due to the fact that conflicts exist between model resource occupation and real-time performance during heat dissipation of the server in the related technology is solved, and the technical effects of optimizing sparse calculation and solving conflicts between multi-mode resources and embedded resources are achieved.

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. Heat dissipation control has become a key challenge to ensure system stability, extend hardware life and reduce energy consumption.

[0003] In the related technology, the server can implement heat dissipation control based on the neural network model, but the neural network model usually requires more memory resources. Since the server heat dissipation needs to meet the real-time control requirements in some scenarios and the precision control requirements in some scenarios, there is a conflict between the model resource occupancy and real-time performance when the server heatsinks. Summary of the invention

[0004] The present invention provides a server heat dissipation control system, method, storage medium and program product, so as to at least solve the technical problems 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 take into account both control accuracy and real-time requirements.

[0005] The present invention provides a server heat dissipation control system, comprising: a data acquisition layer, which acquires multimodal data and system load data of a server; a resource management layer, wherein the resource management layer comprises a storage area pre-allocated with a sparse matrix index table in a memory pool, and a safe area of ​​a neural network model that executes a heat dissipation control task of the server; a data inference layer, which loads the neural network model from the safe area, calculates the current sparsity of the neural network model according to the system load data, queries the sparse matrix index table of the storage area based on the current sparsity, obtains the current weight matrix of the neural network model, and uses the neural network model to infer the multimodal data; and a decision execution layer, which controls the server to execute a heat dissipation action based on a control signal generated by the inference result.

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

[0007] The present invention also provides a computer-readable storage medium, in which a computer program is stored, wherein when the computer program is executed by a processor, the steps of any of the above-mentioned server heat dissipation control methods are implemented.

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

[0009] Through the present invention, the current sparsity of the neural network model is calculated by using the multimodal data and system load data of the server, and then the weight matrix of the neural network model is calculated by the current sparsity, and then the multimodal data is inferred by using the neural network model and the weight matrix, and finally the server is controlled to perform heat dissipation based on the control signal generated by the inference result, and high-performance heat dissipation control is achieved through the technical fusion of dynamic sparse learning and multimodal fusion. As a result, the technical problem that the conflict between model resource occupancy and real-time performance during server heat dissipation in the related technology cannot be solved, resulting in the inability to take into account the requirements of control accuracy and real-time performance. The technical effects of optimizing the sparse calculation of the heat dissipation control technology, resolving the conflict between multimodal and embedded resources, and balancing the learning ability and security are achieved, thereby ensuring the stability of the system, extending the hardware life and reducing energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0011] Figure 1 A schematic diagram of the structure of a server heat dissipation control system provided by an embodiment of the present invention; Figure 2 A server heat dissipation control system architecture diagram provided by an embodiment of the present invention; Figure 3 A flow chart of a server heat dissipation control method provided by an embodiment of the present invention; Figure 4 A schematic diagram of a data collection process provided by an embodiment of the present invention; Figure 5 A schematic diagram of a dynamic model sparse reasoning process provided by an embodiment of the present invention; Figure 6 A schematic diagram of an online incremental learning and security update module provided for one embodiment of the present invention; Figure 7 A schematic flow chart of a method for controlling heat dissipation of a server provided in an embodiment of the present invention.

[0012] Explanation of reference numerals: 101 - data collection layer; 102 - resource management layer; 103 - data reasoning layer; 104 - decision execution layer. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] It should be noted that, in the description of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also 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.

[0015] At present, relevant technologies have many limitations in server heat dissipation control, as follows: (1) Failure of static strategies: Control strategies based on temperature-fan speed mapping tables (such as the IPMI (Intelligent Platform Management Interface) standard) use fixed thresholds and cannot perceive sudden changes in load. For example, when the GPU (Graphics Processing Unit) is fully loaded in AI training, the temperature can soar by 10°C / s. Traditional solutions have the risk of local overheating due to detection-response delays (usually >500ms).

[0016] (2) PID (Proportional-Integral-Derivative) control defects: Although the classic PID algorithm can achieve closed-loop regulation, its linear characteristics make it difficult to cope with the nonlinear thermodynamic system of the server (such as the cubic relationship between fan speed and heat dissipation efficiency). Actual measurement data shows that in the high load mutation scenario, the standard deviation of temperature fluctuation controlled by PID ( ) can reach ±3.5℃, far exceeding industrial requirements ( <2℃).

[0017] (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.

[0018] In AI-driven heat dissipation control technology, the following technical problems exist: (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 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 occupies the BMC (Baseboard Management Controller) security partition space (only 512KB is reserved), which may destroy the firmware integrity verification mechanism.

[0019] (2) Insufficient use of multimodal data: Existing technologies mostly 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 abnormality 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.

[0020] (3) Mismatch between model fixation and dynamic environment: The parameters of the AI ​​model after deployment are fixed and cannot adapt to long-term environmental changes 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). The model needs to be manually retrained, resulting in high operation and maintenance costs.

[0021] 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 the technical problems of the related technologies such as unoptimized sparse computing, multi-sensor timing asynchrony directly affecting the accuracy of control decisions, and conflicts between security and real-time, so as to achieve the technical effects of optimizing sparse computing, resolving conflicts between multi-modality and embedded resources, and balancing learning ability and security, which will be described in detail below.

[0022] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Figure 1 A schematic diagram of a server heat dissipation control system provided by an embodiment of the present invention is shown in FIG. Figure 1As shown, the system includes: a data collection layer 101, a resource management layer 102, a data reasoning layer 103 and a decision execution layer 104.

[0024] Among them, the data acquisition layer 101 is used to collect multimodal data and system load data of the server; the resource management layer 102 includes a storage area with a sparse matrix index table pre-allocated in the memory pool, and a safe area of ​​the neural network model that executes the heat dissipation control task of the server; the data inference layer 103 is used to load the neural network model from the safe area, calculate the current sparsity of the neural network model according to the system load data, query the sparse matrix index table of 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 infer the multimodal data; the decision execution layer 104 is used to control the server to perform heat dissipation actions based on the control signal generated by the inference result.

[0025] The system load data includes computing resource occupancy, memory margin, etc., wherein the computing resource occupancy is the CPU occupancy.

[0026] It can be understood that the embodiment of the present invention includes four main levels: the data acquisition layer 101 is responsible for collecting multimodal data and system load information of the server; the resource management layer 102 has a storage area, in which a sparse matrix index table is pre-allocated and a safe area of ​​the neural network model that performs heat dissipation control tasks is also included; the data reasoning layer 103 loads the neural network model from the safe 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 analyze and reason the multimodal data; finally, the decision execution layer 104 generates a control signal based on the reasoning result to guide the server to perform corresponding heat dissipation measures.

[0027] In an embodiment of the present invention, multimodal data includes sensor data of multiple components in a server, and the data acquisition layer 101 includes a sensor interface and a timer, wherein the sensor interface receives sensor data of multiple components and system load data; the timer generates timestamps for the sensor data of multiple components, and uses the timestamps to align time windows of the sensor data of multiple components.

[0028] Among them, the sensor data of multiple components in the multimodal data server, such as the temperature gradient, power consumption change rate, air flow speed and environmental noise of the CPU, GPU, memory and hard disk, 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 inferred from the current PWM (Pulse Width Modulation) duty cycle to the air flow speed in the computer case, and the environmental noise can be obtained through ADC (Analog-to-Digital Converter); the 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.

[0029] It can be understood that the data acquisition layer 101 of the embodiment of the present invention receives sensor data and system load data from multiple components in the server through the sensor interface, and uses a timer to add an accurate timestamp at the moment each data is acquired. Then, the data of each component is aligned according to the same time window based on these timestamps, ensuring that even if the data sources are different and there may be transmission delays, they can be accurately synchronized to provide support for subsequent comprehensive data analysis and server management.

[0030] In the embodiment of the present invention, the timer is bound to all sensor interrupt sources.

[0031] Among them, the sensor interrupt source is an interrupt request generated to the processor when the sensor detects that the physical quantity it monitors changes and exceeds the preset threshold; associating the hardware timer with all sensors allows the timer to accurately record the time when the event occurs when receiving these interrupt signals, thereby obtaining an accurate timestamp.

[0032] In the embodiment of the present invention, the data acquisition layer 101 includes a preprocessing module and a feature extraction module, wherein the preprocessing module processes multimodal data and system load data; and the feature extraction module extracts features from the preprocessed multimodal data.

[0033] Among them, the preprocessing module is used to preprocess the original multimodal data and system load data obtained from sensors and system monitoring tools, including normalization, filtering (such as sliding average denoising) and other operations to improve data quality and prepare for subsequent analysis; the feature extraction module extracts features from the preprocessed multimodal data, such as calculating the temperature change rate.

[0034] It can be understood that in the data acquisition layer 101 of the embodiment of the present invention, a series of preprocessing operations such as normalization and filtering are performed on the multimodal data and system load data through the preprocessing module to improve the data quality and consistency; the feature extraction module is used to extract features from the preprocessed multimodal data to extract features such as the temperature change rate, thereby ensuring the effective conversion from raw data to useful information and enhancing the intelligence level and response speed of the system.

[0035] In an embodiment of the present invention, the secure area stores an encrypted neural network model and is used for security verification of the neural network model.

[0036] Among them, the secure area is an isolated environment that can divide the processor's working status 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 encryption of the trained neural network model to prevent unauthorized access or tampering. The encrypted model can only be loaded and used after the correct decryption process.

[0037] It can be understood that the embodiment of the present invention adopts a secure area, in which an encrypted neural network model is stored, and is used to perform 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 loaded and used normally; if the verification fails, the model loading is prevented, thereby ensuring the security and reliability of the system.

[0038] In an embodiment of the present invention, the security verification process 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 a compressed sparse column format and storing it in the target address of the memory pool.

[0039] Among them, the hash value is a fixed-length string generated by a hash function for the encrypted neural network model. The string can be used as a data fingerprint to verify the integrity and consistency of the data; the pre-stored value is a 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, namely CSC (Compressed Sparse Column), is a method of storing sparse matrices that only saves non-zero elements and their corresponding row indices and column pointers to reduce memory usage and speed up specific operations.

[0040] It can be understood that in the security verification process of the embodiment of the present invention, the hash value of the encrypted neural network model is first 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 means 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 storage and computing 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, thereby enhancing the security and performance of the overall system.

[0041] In an embodiment of the present invention, the data inference layer 103 includes an adjustment module, wherein 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.

[0042] Among them, sparsity refers to the proportion of non-zero parameters in the neural network model. Higher sparsity means that more parameters are set to zero, thereby reducing the amount of calculation and storage requirements. Sparsity is usually expressed as a value between 0 and 1, where 0 means completely dense and 1 means completely sparse. The following formula can be used to calculate the current sparsity based on the computing resource occupancy rate: , is the current sparsity, is the sigmoid function, which is used to map the x value to the (0,1) interval to achieve a smooth transition. .

[0043] It can be understood that the embodiment of the present invention first loads the decrypted secure neural network model through the adjustment module, 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 positively proportional to the computing resource occupancy rate, that is, the higher the computing resource occupancy rate, the higher the sparsity of the neural network model. Conversely, the lower the computing resource occupancy rate, the lower the sparsity of the neural network model. For example, when the computing resource occupancy rate is high, the server is under high load. At this time, the calculated sparsity is 90%, which sacrifices part of the accuracy to ensure real-time performance, thereby achieving a trade-off between accuracy and delay. When the computing resource occupancy rate is low, the server is under low load. At this time, the calculated sparsity is 30%, which appropriately increases the computing amount and storage requirements, thereby improving the model accuracy and optimizing resource utilization.

[0044] In an embodiment of the present invention, the data inference layer 103 includes a computing module, wherein the computing module converts the current weight matrix into a compressed sparse column format and implements parallel multiplication and addition operations of the current weight matrix using a target instruction set.

[0045] The target instruction set refers to the instruction set supported by a specific hardware platform, such as the NEON SIMD (single instruction multiple data stream) extension in the ARM architecture. These instruction sets can accelerate specific types of computing operations, such as matrix multiplication.

[0046] It can be understood that in the data inference layer 103 of the embodiment of the present invention, the computing module first converts the current weight matrix of the neural network model into a compressed sparse column format to optimize storage and computing efficiency; then, the instruction set provided by the target hardware platform is used to implement parallel multiplication and addition operations on the converted weight matrix, thereby improving the speed of matrix operations and reducing energy consumption by efficiently utilizing hardware resources, thereby ensuring high-performance operation of the system.

[0047] In the embodiment of the present invention, the data reasoning layer 103 includes a fusion module, wherein the fusion module fuses the multimodal data into a multidimensional feature vector, inputs the multidimensional feature vector into a neural network model, and the neural network model outputs a reasoning result.

[0048] The multimodal data are fused into a multidimensional feature vector. For example, the temperature gradient, power consumption change rate, and air flow velocity in the multimodal data are spliced ​​into a three-dimensional feature vector.

[0049] It can be understood that in the data reasoning layer 103 of the embodiment of the present invention, a fusion module is used to integrate multimodal data into a multidimensional feature vector, and then the multidimensional feature vector is sent to the neural network model, which is used to perform in-depth analysis on it and generate corresponding reasoning results, so that the system can optimize the heat dissipation control strategy based on accurate predictions to ensure efficient and stable operation.

[0050] In an embodiment of the present invention, the decision execution layer 104 includes an optimization control module, a hardware driving module and a feedback adjustment module, wherein the optimization control module takes multiple of temperature stability, fan life and system energy efficiency as optimization targets, generates a first driving signal according to a control signal and an error signal fed back by the feedback adjustment module, and sends the first driving signal to the hardware driving module; the hardware driving module drives the heat dissipation array to perform heat dissipation action based on the first driving signal.

[0051] The heat dissipation array refers to a group of fans or other cooling devices inside the server for heat dissipation.

[0052] It can be understood that in the decision execution layer 104 of the embodiment 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, combined with the reasoning result and the error signal provided by the feedback adjustment module. The hardware driver module executes specific heat dissipation measures according to the drive signal by directly controlling the action of the heat dissipation array, thereby 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, thereby realizing intelligent and efficient server heat dissipation management.

[0053] In an embodiment of the present invention, the decision execution layer 104 also includes an abnormal fuse mechanism, wherein the abnormal fuse mechanism includes: obtaining the current temperature of the server; if the current temperature exceeds the temperature threshold, generating a second drive signal; sending the second drive signal to the hardware drive module, and the hardware drive module drives the heat dissipation array to perform heat dissipation action based on the second drive signal.

[0054] Among them, the temperature threshold is a pre-set safety temperature upper limit. When the actual temperature of the server exceeds the temperature threshold, the corresponding heat dissipation action will be triggered. The temperature threshold is set according to the actual situation and is not specifically limited here.

[0055] It can be understood that in the embodiment of the present invention, in the decision execution layer 104, the current temperature of the server is monitored in real time through the abnormal fuse mechanism, that is, once it is detected that the temperature exceeds the preset temperature threshold, a second drive signal is immediately generated, and the second drive signal is sent to the hardware driver module, prompting the heat dissipation array to quickly perform enhanced heat dissipation actions, thereby achieving a rapid response when the server temperature rises abnormally, performing enhanced heat dissipation actions, effectively preventing hardware from being damaged due to overheating, and ensuring the stability and security of the system.

[0056] In the embodiment of the present invention, the resource management layer 102 responds to the server heat dissipation control demand and starts the server heat dissipation control thread, which is used to execute the server heat dissipation control task and set the priority of the heat dissipation control task.

[0057] 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 first, ensuring that once temperature anomalies or other heat dissipation needs are detected, immediate action can be taken.

[0058] It is understandable that in order to ensure that the heat dissipation action can respond quickly when necessary and avoid hardware damage or other performance problems due to overheating, the embodiment of the present invention will start a special heat dissipation control thread at the resource management layer 102, which is responsible for performing 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 be processed in a timely manner under high load conditions, thereby ensuring that the server can maintain good heat dissipation under various working conditions, avoiding the risks brought by overheating, ensuring the stability and reliability of the system, and through the reasonable allocation of resources and task priorities, the system can efficiently respond to real-time heat dissipation needs and maintain the best operating state.

[0059] 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.

[0060] Among them, the interrupt response threshold needs to match the fan Tachometer signal (Tachometer signal is a pulse signal output by a sensor built into the fan, which is proportional to the rotation speed of the fan) feedback requirement, and is set according to actual needs. No specific limitation is made here. Binding the interrupt to the target processing core of the server can reduce cross-core switching delay.

[0061] It can be understood that in order to match the fan Tachometer signal feedback requirements, the embodiment of the present invention needs to identify the Tachometer signal characteristics provided by the fan before starting the server cooling control thread, and set the interrupt response threshold accordingly, bind the interrupt to the target processing core of the server, and reduce the cross-core switching delay.

[0062] In the embodiment of the present invention, the resource management layer 102 configures a direct memory access channel before starting the heat dissipation control thread of the server, and uses the direct memory access channel to transmit the sensor data to the hardware acceleration engine corresponding to the sensor interface.

[0063] Among them, the direct memory access channel hardware acceleration engine refers to the part used to accelerate sensor data analysis and processing.

[0064] It can be understood that before starting the heat dissipation control thread of the server, the embodiment of the present invention configures a direct memory access channel at the resource management layer 102, and utilizes the direct memory access channel to realize direct transmission of sensor data to the hardware acceleration engine corresponding to the sensor interface, thereby avoiding the CPU as an intermediary, so that the sensor data can 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, ensuring the stability and efficient operation of the system.

[0065] In an embodiment of the present invention, the server heat dissipation control system also includes: a model update layer, wherein the model update layer includes a monitoring module, an update module and a safe hot switching module, the monitoring module obtains monitoring data of the server during the heat dissipation control process, and predicts the performance degradation rate of the neural network model based on the monitoring data; the update module performs a 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 safe hot switching module implements switching between the neural network model before the update and the neural network model after the update based on integrity verification; the update module clips the updated gradient when updating to limit the update amplitude.

[0066] Among them, monitoring data refers to the data collected from various components of the server during the heat dissipation control process, including temperature, power consumption, airflow speed, fan speed, etc., which are used to evaluate the current heat dissipation status and system performance; performance decay rate: measures the degree to which the prediction accuracy and effect of the model decrease over time. It is usually calculated by monitoring the difference between the actual running data and the model prediction results; gradient clipping can avoid the gradient explosion problem by limiting the maximum value of the gradient, thereby maintaining the stability of model training; the update threshold is set according to actual needs and is not specifically limited here.

[0067] 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 data. When the performance degradation rate exceeds the set update threshold, the update module starts the parameter update process, during which the gradient clipping technology is used to limit the update amplitude. At the same time, the safe hot switching module ensures the correctness and security of the new model based on the integrity check, and realizes seamless switching between the old and new models after 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, so as to maintain efficient and stable heat dissipation control performance.

[0068] According to the server heat dissipation control system proposed in the embodiment of the present invention, through the data collection layer, resource management layer, data inference layer and decision execution layer, it is possible to use the multimodal data and system load data of the server to calculate the current sparsity of the neural network model, and then calculate the weight matrix of the neural network model through the current sparsity, and then use the neural network model and the weight matrix to infer the multimodal data. Finally, based on the control signal generated by the inference result, the server is controlled to perform the heat dissipation action. Through the technical fusion of dynamic sparse learning and multimodal fusion, a lightweight timing alignment algorithm is designed, a secure learning container is constructed, and high-performance heat dissipation control is realized, achieving the technical effects of optimizing sparse computing in heat dissipation control technology, resolving conflicts between multimodality and embedded resources, and balancing learning ability and security, ensuring the stability of the system, extending the hardware life and reducing energy consumption.

[0069] The server heat dissipation control system is further described below through a specific embodiment.

[0070] The server heat dissipation control system architecture of this embodiment is as follows Figure 2 As shown in the figure, it is mainly composed of a multimodal 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, as follows: 1. Multimodal Data Collection Layer Heterogeneous sensor data collection: Real-time acquisition of data such as the temperature (obtained through IPMI), power consumption (obtained through PMBus), chassis airflow speed (obtained through PWM reverse calculation), and environmental noise (obtained through ADC) of each server component (CPU, GPU, memory, hard disk).

[0071] Hardware-level time alignment: Use the built-in hardware timer (accuracy 1μs) to timestamp all sensor data and solve the asynchronous sampling problem (error <1ms).

[0072] Data preprocessing: normalize, filter (such as sliding average denoising) and extract features (such as calculating the temperature change rate dT / dt) of the raw data.

[0073] 2. Dynamic Sparse Learning Engine Variable sparsity inference: Dynamically adjust the sparsity of the neural network (30%-90%) according to the real-time system load (CPU usage, memory margin) to balance calculation accuracy and latency.

[0074] Sparse matrix accelerated computing: Convert model weights into a compressed sparse column format, and use instruction sets (such as the AST2600 NEON SIMD instruction set) to implement parallel multiplication and addition operations, improving sparse computing efficiency by 3.2 times.

[0075] Multimodal feature fusion: The temperature gradient, power consumption change rate, and airflow velocity are concatenated into a three-dimensional feature vector and input into a dynamic sparse network for joint reasoning.

[0076] 3. Real-time decision execution layer Multi-objective optimization control: Generates PWM control signals based on inference results, while optimizing temperature stability (σ<2°C), fan life (speed change rate <200rpm / s), and system energy efficiency (PUE<1.1).

[0077] Hardware PWM direct drive: The hardware PWM controller directly outputs signals to the fan array, with an end-to-end delay of less than 1ms.

[0078] Abnormal fuse mechanism: When the temperature exceeds the limit (such as >95°C), the algorithm is bypassed and the full-speed cooling mode is directly triggered.

[0079] 4. Online Incremental Learning Module Sliding window trigger: Continuously monitor the control effect (such as temperature fluctuation and energy consumption) and activate learning when the performance decay within the window (30 seconds) exceeds the threshold.

[0080] Selective parameter update: Only the gradients of key neuron weights (<5%) are updated, reducing the amount of calculation by 90%.

[0081] Secure model hot switching: Double-buffered memory design is used to achieve seamless switching between old and new models, combined with TrustZone (secure area) to encrypt and store model parameters.

[0082] 5. Hardware Resource Management Layer Memory pool pre-allocation: Pre-allocate sparse matrix index table and non-zero weight storage area in SRAM (Static Random Access Memory) to avoid DDR4 (Double Data Rate 4) access latency.

[0083] Task priority scheduling: The heat dissipation control task is set to the highest real-time priority (priority = 0), which can preempt other BMC background services (such as logging).

[0084] Figure 3 The flowchart of the server heat dissipation control method provided in this embodiment is as follows: Figure 3 As shown, the method comprises the following steps: Step 1: System initialization phase: 1. Hardware resource configuration SRAM pre-allocation: SRAM dedicated area: 512KB, storing sparse indexes (120KB, CSC format) and weights (360KB, FP16 (Half Precision Floating Point) quantization), 32KB redundancy is reserved, and the dynamic memory usage of the model during runtime does not exceed the reserved redundant space (32KB).

[0085] DDR4 dual buffer pool: pre-allocated 1MB area for redundant backup during model hot switching.

[0086] 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 used to accelerate the SIMD architecture extension of multimedia and signal processing applications; SIMD, or single instruction, multiple data, is a parallel computing technology that allows one instruction to process multiple data points simultaneously, thereby improving the efficiency of specific types of computing tasks) extension of the processor core, set the floating-point mode to half precision (FP16), and adapt the quantization weight format of the dynamic sparse model.

[0087] Configure the DMA (Direct Memory Access) channel to achieve zero-copy transmission of sensor data to the hardware acceleration engine.

[0088] Enable the TrustZone security area. The TrustZone security area is designed to provide a hardware-level security environment for the device. It protects sensitive data and code by creating two independent execution areas at the processor level, namely the secure area and the non-secure area. It ensures strict isolation between the secure area and the non-secure area, and divides the secure area (storing encryption models) and the non-secure area (processing regular tasks).

[0089] Peripheral interface initialization: configure IPMI interface, enable PMBus interface, initialize PWM controller, and start ADC channel.

[0090] Timer synchronization: Configure the TIM3 (Timer 3) hardware timer and bind it to all sensor interrupt sources.

[0091] The hardware resource configuration design is highly compatible and embedded-friendly. The model memory occupies less than 512KB, adapts to the DDR4 memory limit, and avoids memory fragmentation problems through SRAM pre-allocation. It can run continuously for 72 hours without memory leaks. The security enhancement mechanism is based on TrustZone to achieve model encryption storage and isolation from incremental learning processes to prevent malicious code injection attacks. The CRC32 hardware engine verifies the integrity of the model and has a low bit error rate.

[0092] 2. Model loading and verification Load a pre-trained dynamic sparse model (initial sparsity 70%, number of parameters <100K).

[0093] Use the CRC32 (Cyclic Redundancy Check 32) hardware engine.

[0094] Calculate the model hash value and compare it with the pre-stored value to complete the integrity check.

[0095] Convert the model weights into CSC format and store them in the specified SRAM address (0x1E720000).

[0096] Through the coordinated optimization of the dynamic sparse learning engine and hardware acceleration, the end-to-end control delay is <15ms (when the sparsity is >60%), which is a significant improvement over traditional solutions and can effectively cope with GPU / CPU transient load mutations (such as a power surge of 200W / s). Fast multi-modal data fusion: Use hardware timers to achieve multi-sensor data time alignment (error <1ms), eliminate the 10ms-level error caused by traditional software interpolation, and increase the lead time of anomaly detection to more than 200ms.

[0097] Step 2: Real-time task scheduling and interrupt optimization 1. Real-time kernel configuration Apply the PREEMPT_RT (Real-Time Preemption Patch) patch to the Linux kernel to enable fully preemptive scheduling.

[0098] Set the priority of the heat dissipation control thread to ensure the highest real-time performance: 2. Interrupt Binding and Optimization Bind the TIM3 interrupt (IRQ (Interrupt Request) 45) to CPU Core 0 to reduce cross-core switching delay: Set the PWM controller interrupt response threshold to 10μs to match the fan Tachometer signal feedback requirement.

[0099] 3. The main control thread starts Create a real-time control thread and loop to wait for the TIM3 interrupt to be triggered (1ms period).

[0100] Step 3: Multimodal data acquisition and synchronization The data collection process is as follows Figure 4 As shown, the details are as follows: 1. Sensor data reading Temperature: Read the CPU temperature value through IPMI instructions.

[0101] Power consumption: PMBus command to read instantaneous power.

[0102] Airflow speed: Based on the current PWM duty cycle, the CMF (Cubic Feet per Minute, i.e. airflow speed) value is inferred: ,in, is the PWM duty cycle, that is, the proportion of the high level time in the PWM signal to the entire cycle.

[0103] Noise analysis: After ADC sampling, a 256-point FFT (Fast Fourier Transform) is performed using a Hamming window to extract the energy in the 50Hz-1kHz frequency band.

[0104] 2. Hardware timestamp synchronization Each time data arrives, the TIM3 counter value is read (1μs accuracy) and stored in a structure with a timestamp.

[0105] Time window alignment: Use 1ms as the window and discard packets with a time deviation exceeding ±0.5ms.

[0106] 3. Data buffering and filtering Double-buffered queue: Uses a lock-free ring buffer to store aligned data.

[0107] Sliding average filtering: A 5-point window filter is applied to the temperature data, and exponential smoothing (α=0.2) is used for the power consumption data.

[0108] The multimodal data are fused into a multidimensional feature vector. For example, the temperature gradient, power consumption change rate, and airflow velocity in the multimodal data are spliced ​​into a three-dimensional feature vector.

[0109] Step 4: Dynamic Sparse Model Inference The specific reasoning flow chart is as follows Figure 5 As shown, specifically including: 1. Dynamic adjustment of sparsity:

[0110] The CPU utilization is read in real time through the PMU counter to calculate the sparsity. Ucpu is the real-time CPU utilization percentage (0-100%), which needs to be normalized before input. The sigmoid function is used to map the x value to the (0,1) interval to achieve a smooth transition. .

[0111] 2. Sparse matrix loading: Load the pre-trained CSC format model according to the current sparsity: Column pointer array (col_ptr): 32-bit integer that records the starting position of non-zero elements in each column.

[0112] Row index array (row_ind): 16-bit integer, compressed storage of row numbers.

[0113] Non-zero weights (values): FP16 format, loaded into SRAM with 4KB alignment.

[0114] 3.NEON instructions accelerate computing Use ARM assembly to optimize sparse matrix multiplication, such as using vld1.16 to load weights, Vmla.f16 multiplication and addition operations, and loop unrolling x4 optimization to achieve NEON instruction accelerated computing.

[0115] Step 5: Control signal generation and closed-loop feedback 1. Multi-objective optimization solution Minimize the cost function:

[0116] Among them, σT is the standard deviation of temperature, reflecting temperature fluctuations. Reducing σT can improve system stability. Δrpm is the speed change, indicating the adjustment range of fan speed. Reducing Δrpm can extend the life of equipment. PUE (Power Usage Effectiveness) is the energy efficiency indicator of data centers, which is the ratio of total energy consumption to IT equipment energy consumption. The closer the PUE is to 1, the higher the energy efficiency.

[0117] Gradient descent iteration: Iterate 3 times with a fixed step size (0.01) to find the optimal PWM duty cycle.

[0118] 2.PWM signal output Write the calculated duty cycle into the PWM controller register.

[0119] 3. Closed loop speed regulation Read the fan Tachometer signal and calculate the speed error: ,in, is the target speed, i.e. the expected speed set by the system according to demand; is the actual speed, that is, the current speed of the fan measured in real time through the Tachometer signal; The speed error is the difference between the target speed and the actual speed, reflecting the speed deviation that the system needs to compensate.

[0120] PID regulation.

[0121] Through the PID control algorithm, the control output is adjusted in combination with the speed error to achieve the desired speed, realize closed-loop feedback regulation, and ensure that the system can respond dynamically and stabilize in the target state.

[0122] Through multi-objective optimization strategies (temperature stability, fan life, and system energy efficiency joint optimization), under the same heat dissipation requirements, fan energy consumption is reduced compared to traditional solutions. Sparse computing energy efficiency ratio is improved: NEON instructions accelerate sparse matrix operations, making the computing power per unit power consumption reach 3.2 times that of traditional intensive computing, which is suitable for low-power embedded scenarios.

[0123] Step 6: Online incremental learning and security updates The process of online incremental learning and security update is as follows: Figure 6 As shown, specifically including: 1. Performance monitoring trigger Calculate the temperature prediction RMSE (Root Mean Square Error) every 30 seconds:

[0124] Among them, N=3000 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.

[0125] If the RMSE exceeds the baseline by 5%, incremental learning is triggered.

[0126] 2. Lightweight parameter update Gradient mask generation: After back-propagation, select the parameters with the top 5% absolute gradient values ​​(sorted by L1 norm).

[0127] Gradient clipping: Limit the update amplitude (Δw ∈ [-0.1, 0.1]) to prevent overfitting.

[0128] Generate an incremental parameter package and encrypt the parameter package.

[0129] 3. Safe hot switching Double buffer switching: Atomic operation switches model pointers (the old model reserves 3 cycles for rollback).

[0130] Integrity check: Calculate the model hash (polynomial 0x04C11DB7), which takes 0.8ms. If the check fails, roll back the old model, issue an alarm, and record the SEL (System Event Log) event. If the check passes, switch the model pointer and update the model.

[0131] Step 7: Exception handling and circuit breaking 1. Temperature exceeds the limit and fuses out Hardware comparator: real-time monitoring of temperature signals, trigger conditions: 85°C: IPMI SEL events are recorded and alarms are triggered.

[0132] 95℃: Forced PWM = 100% (full speed cooling).

[0133] 105°C: Pull down the GPIO (General-Purpose Input / Output) 12 pin and cut off the power supply through the CPLD (Complex Programmable Logic Device) (delay <3ms).

[0134] 2. Safety recovery mechanism After the fuse is blown, the temperature must drop below 75°C (80°C - 5°C hysteresis) before switching back to the normal mode.

[0135] After the fuse signal is triggered, it remains at a low level for at least 10ms to avoid accidental power failure caused by transient temperature fluctuations.

[0136] Through online incremental learning, model parameters can be updated online (only the top 5% key weights are updated), and long-term changes such as hardware aging and changes in the layout of the computer room can be adapted without downtime. The performance degradation rate after the model update is <1%. Dynamic sparsity adjustment: The model sparsity (30%-90%) is adjusted in real time according to the system load, and latency is prioritized under high load (sparseness 90%), and accuracy is improved under low load (sparseness 30%), so as to achieve the optimal allocation of computing resources and make the model have long-term environmental adaptability.

[0137] Figure 7 A schematic flow chart of a server heat dissipation control method provided by an embodiment of the present invention is provided. The method performs heat dissipation control based on the above-mentioned server heat dissipation control system. Figure 7 As shown, the method specifically comprises the following steps: In step S201, multimodal data and system load data of the server are obtained.

[0138] It can be understood that the embodiment of the present invention first needs to obtain multimodal data composed of temperature gradients, power consumption change rates, air flow speeds, and environmental noise of various server components such as the CPU, GPU, memory, and hard disk through various sensors or interface units, and query system load data such as CPU occupancy, memory remaining, etc.

[0139] In step S202, the current sparsity of the neural network model is calculated according to the system load data.

[0140] Among them, sparsity refers to the proportion of non-zero parameters in the neural network model. Higher sparsity means that more parameters are set to zero, thereby reducing the amount of calculation and storage requirements. The current sparsity of the neural network model is calculated according to the system load data, including: identifying the computing resource occupancy rate in the system load data; calculating the current sparsity according to the computing resource occupancy rate.

[0141] It can be understood that the current sparsity of the embodiment of the present invention is obtained by calculating the computing resource occupancy rate, and 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 direct proportion to the computing resource occupancy rate, that is, the higher the computing resource occupancy rate, the higher the sparsity of the neural network model. Conversely, the lower the computing resource occupancy rate, the lower the sparsity of the neural network model. For example, when the computing resource occupancy rate is high, the server is under high load, and the sparsity calculated at this time is 90%, which sacrifices part of the accuracy to ensure real-time performance, thereby achieving a trade-off between accuracy and delay. When the computing resource occupancy rate is low, the server is under low load, and the sparsity calculated at this time is 30%, which appropriately increases the amount of calculation and storage requirements, thereby improving the model accuracy.

[0142] In step S203, the weight matrix of the neural network model is calculated based on the current sparsity, and the multimodal data is inferred using the neural network model and the weight matrix.

[0143] Among them, the weight matrix is ​​a collection of connection weights between neural network layers, which represents the connection strength between each neuron; the neural network model is sparsed using the current sparsity. During the sparsification process, many elements in the weight matrix will be set to zero to form a sparse matrix; the neural network model and the weight matrix are used to infer multimodal data, including: extracting multimodal features from multimodal data and fusing multimodal features; inputting the fused multimodal features into the neural network model, and the neural network model outputs the inference results, and the weight matrix is ​​used to accelerate the inference process of the neural network model.

[0144] It can be understood that the embodiment of the present invention utilizes a neural network model and its corresponding sparse weight matrix to infer multimodal data. First, it is necessary to extract the obtained multimodal data, fuse the extracted features of various types, input the fused multimodal feature vector into the neural network model with the current sparsity dynamically adjusted, and output the inference result through the neural network model, thereby reducing the computational complexity and ensuring the accuracy and real-time performance of the inference.

[0145] In step S204, based on the control signal generated by the inference result, the server is controlled to perform a heat dissipation action.

[0146] It can be understood that the embodiment of the present invention infers the current thermal state of the server through reasoning, and generates a specific control signal based on the reasoning 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 signal and controls the fan speed, thereby effectively managing the internal temperature and ensuring that the hardware runs under optimal working conditions.

[0147] For the description of the features in the embodiment corresponding to the server heat dissipation control method, reference can be made to the relevant description of the embodiment corresponding to the server heat dissipation control system, which will not be repeated here.

[0148] According to the server heat dissipation control method proposed in the embodiment of the present invention, the current sparsity of the neural network model is calculated by using the multimodal data and system load data of the server, and then the weight matrix of the neural network model is calculated by 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 heat dissipation actions. Through the technical fusion of dynamic sparse learning and multimodal fusion, a lightweight timing alignment algorithm is designed, a secure learning container is constructed, and high-performance heat dissipation control is realized, achieving the technical effects of optimizing sparse computing in heat dissipation control technology, resolving conflicts between multimodal and embedded resources, and balancing learning ability and security, thereby ensuring the stability of the system, extending the hardware life, and reducing energy consumption.

[0149] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method.

[0150] An embodiment of the present invention further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above-mentioned server heat dissipation control method embodiments when running.

[0151] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0152] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above-mentioned server heat dissipation control method embodiments are implemented.

[0153] An embodiment of the present invention further provides another computer program product, including a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in any of the above-mentioned server heat dissipation control method embodiments are implemented.

[0154] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0155] The above is a detailed introduction to a server heat dissipation control method, electronic device, storage medium and program product provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A server heat dissipation control system, characterized in that: include: Data collection layer, collects multi-modal data and system load data of the server; A resource management layer, wherein the resource management layer includes a storage area pre-allocated with a sparse matrix index table in a memory pool and a security area of ​​a neural network model that performs a heat dissipation control task of the server; A data inference layer, which loads the neural network model from the secure area, calculates the current sparsity of the neural network model according to the system load data, queries the sparse matrix index table of 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 infer the multimodal data; The decision execution layer controls the server to perform heat dissipation actions based on the control signal generated by the inference result.

2. The server heat dissipation control system according to claim 1, characterized in that: The multimodal data includes sensor data of multiple components in the server, and the data acquisition layer includes a sensor interface and a timer, wherein: The sensor interface receives sensor data of the plurality of components and the system load data; The timer generates a timestamp for the sensor data of the multiple components, and uses the timestamp to align time windows of the sensor data of the multiple components.

3. The server heat dissipation control system according to claim 2, characterized in that: 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, wherein: The preprocessing module preprocesses the multimodal data and the system load data; The feature extraction module extracts features from the preprocessed multimodal data.

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

6. The server heat dissipation control system according to claim 5, characterized in that: The safety verification process includes: Calculating a 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 check is passed, the initial weight matrix of the neural network model is converted into a compressed sparse column format and stored in the target address of the memory pool.

7. The server heat dissipation control system according to claim 5, characterized in that: The data inference layer includes an adjustment module, wherein the adjustment module loads the 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.

8. The server heat dissipation control system according to claim 7, characterized in that: The data inference layer includes a computing module, wherein the computing module converts the current weight matrix into a compressed sparse column format and implements parallel multiplication and addition operations of the current weight matrix using a target instruction set.

9. The server heat dissipation control system according to claim 8, characterized in that: The data inference layer includes a fusion module, wherein the fusion module fuses the multimodal data into a multidimensional feature vector, inputs the multidimensional feature vector into the neural network model, and the neural network model outputs an inference result.

10. The server heat dissipation control system according to claim 1, characterized in that: The decision execution layer includes an optimization control module, a hardware drive module and a feedback adjustment module, wherein: The optimization control module takes multiple of temperature stability, fan life and system energy efficiency as optimization targets, 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 drive module; The hardware driving module drives the heat dissipation array to perform a heat dissipation action based on the first driving signal.

11. The server heat dissipation control system according to claim 10, characterized in that: The decision execution layer also includes an abnormal fuse mechanism, wherein the abnormal fuse mechanism includes: Obtaining the current temperature of the server; If the current temperature exceeds the temperature threshold, generating a second driving signal; The second driving signal is sent 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.

12. The server heat dissipation control system according to claim 1, characterized in that: The resource management layer responds to the heat dissipation control demand of the server, 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.

13. 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 to 12, wherein the method comprises: Obtain multimodal data and system load data of the server; Loading the neural network model from the safe area, calculating the current sparsity of the neural network model according to the system load data, querying the sparse matrix index table of 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 infer the multimodal data; Based on the control signal generated by the inference result, the server is controlled to perform a heat dissipation action.

14. 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 server heat dissipation control method as claimed in claim 13.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the server heat dissipation control method as claimed in claim 13 are implemented.

Images