A server heat dissipation regulation method and device and medium

Abstract

This invention relates to the field of heat dissipation control technology, specifically to a server heat dissipation control method, device, and medium. The method includes: acquiring core temperature, cumulative error, and a task vector containing characteristics such as instruction structure and floating-point density; predicting the expected heat flux based on a model, calculating a thermal penalty factor positively correlated with the cumulative error, and generating a scheduling scheme with the goal of balancing the penalized heat flux with the total execution time; calculating the cumulative error by setting an integral step size inversely proportional to the rate of change of temperature error; predicting the temperature trajectory using a first-order heat conduction model with the initial slope as the differential term; calculating the proportional gain relating the expected heat flux and the cumulative error, and summing the control variables to generate a heat dissipation control command. The solution of this invention can avoid local hotspots, enhance control predictability, and achieve stable and coordinated temperature control.

Description

Technical Field

[0001] This invention relates to the field of heat dissipation control technology. More specifically, this invention relates to a server heat dissipation control method, apparatus, and medium. Background Technology

[0002] With the rapid development of server hardware technology and the continuous improvement of chip integration, the power density of core computing units such as CPUs and GPUs has increased dramatically. This high power density operation results in servers generating enormous heat when handling heavy workloads. If heat is not dissipated effectively and promptly, it will seriously affect the stability and lifespan of the equipment. To maintain normal system operation, mainstream server cooling solutions typically employ a temperature sensor-based feedback control mechanism. Specifically, this mechanism monitors the temperature of the core area in real time. When the detected temperature exceeds a preset safety threshold, the system controls the cooling fans to increase their speed, aiming to achieve a cooling effect by enhancing air convection, thereby preventing hardware overheating.

[0003] To further optimize the aforementioned heat dissipation control effect, existing technologies often incorporate proportional-integral-derivative (PID) control algorithms into the cooling system. The PID control algorithm not only considers the current temperature error but also combines historical accumulated errors and predictions of future error trends for comprehensive calculations. Compared to simple threshold switching control, the PID algorithm can output a more precise control signal, thus providing a smoother and more timely response. This improved control method effectively suppresses drastic temperature fluctuations to a certain extent, enhances the dynamic adjustment capability of the cooling system to temperature changes, and enables the server to maintain a relatively stable operating temperature within a certain range.

[0004] However, the aforementioned existing technologies still have significant shortcomings in practical applications. First, the control logic of traditional threshold control and PID control is based on post-event temperature measurements, exhibiting significant lag and failing to predict and intervene before temperature rises. Second, at the task scheduling level, existing schedulers typically only optimize computational performance indicators such as throughput and latency, ignoring thermal power consumption. This single-dimensional scheduling strategy may unintentionally concentrate multiple high-thermal-power tasks on a few cores, causing these cores to form localized hotspots that are difficult to eliminate with conventional fan adjustments within a short period. Furthermore, the differential term of the PID algorithm is calculated based on historical data, making it highly sensitive to latency and easily affected by signal noise; while the integral term is prone to integral saturation when errors accumulate over a long period, leading to overshoot and oscillations, which in turn undermines system stability. In summary, existing technologies lack a perspective that deeply integrates task scheduling and thermal control, and also lack advanced strategies that can adapt to dynamic load changes and accurately predict future thermal trends, making it difficult to achieve low-energy, high-stability coordinated regulation of server temperature. Summary of the Invention

[0005] The purpose of this invention is to propose a server heat dissipation control method, device, and medium to solve the problem of difficulty in achieving low-energy-consumption and high-stability coordinated control of server temperature in the prior art; to this end, this invention provides solutions in the following four aspects.

[0006] In a first aspect, the present invention provides a server heat dissipation control method, comprising the following steps:

[0007] The current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled are obtained. The task feature vector includes the instruction type composition ratio, floating-point operation density and memory access frequency of the task.

[0008] Based on the task feature vector and the preset instruction set thermal power consumption model, predict the expected heat flux time series of the task set to be scheduled on each core; calculate the thermal penalty factor that is positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle; generate and execute a task scheduling scheme with the goal of balancing the expected heat flux of each core and the total execution time, which are included in the thermal penalty factor.

[0009] Calculate the temperature error between the current temperature and the target temperature; set an integral time step that is inversely proportional to the absolute value of the rate of change of temperature error, and calculate the cumulative temperature error integral value of the current period using the trapezoidal integral method; use the expected heat flux time series and the current temperature to predict the temperature change trajectory within the future time window through the server's first-order heat conduction model, and use the initial slope of the trajectory as the differential term.

[0010] Calculate a proportional gain coefficient that is positively correlated with the absolute value of the expected heat flux and the cumulative temperature error integral value of the current cycle; based on the proportional gain coefficient, the cumulative temperature error integral value and the differential term, generate proportional, integral and differential control quantities and sum them to generate a heat dissipation control command for adjusting the speed of the server cooling fan.

[0011] Preferably, the step of predicting the expected heat flux time series of the task set to be scheduled executed on each core includes: decomposing each task in the task set into a sequence of micro-operations within a unit time slice; querying a preset instruction set heat power model to obtain the basic heat power value corresponding to each micro-operation; weighting and summing the basic heat power values ​​of each micro-operation according to the instruction type composition ratio, floating-point operation density and memory access frequency in the task feature vector to obtain the instantaneous heat power within the time slice; and concatenating the instantaneous heat power values ​​in consecutive time slices to obtain the expected heat flux time series executed on the corresponding core.

[0012] Preferably, the formula for calculating the thermal penalty factor is: ,in, For the first One core thermal penalty factor, For the first The cumulative temperature error integral value of the previous cycle of each core. This is the preset penalty gain coefficient.

[0013] Preferably, the step of generating and executing a task scheduling scheme with the goal of balancing the expected heat flux of each core component and the total execution time, which are included in the thermal penalty factor, includes:

[0014] Construct the optimization objective function ,in, The total completion time for the task set. For the first One core thermal penalty factor, To be allocated to the first The expected total heat production of all tasks in each core, and For preset weighting coefficients, and These are the normalized time and normalized heat used to make the components of the objective function dimensionless; the particle swarm optimization algorithm is used to iteratively optimize the allocation scheme of tasks on each core until the objective function is achieved. The value of converges to the minimum, thus obtaining the task scheduling scheme.

[0015] Preferably, the step of calculating the cumulative temperature error integral value for the current period includes:

[0016] Calculate the current temperature error Temperature error compared to the previous cycle The difference is used to obtain the change in temperature error. ;

[0017] According to the formula Calculate the integration time step ,in, Based on the time step, This is the preset integral step size adjustment coefficient;

[0018] Through formula Calculate the cumulative temperature error integral value for the current period. ,in, This is the cumulative temperature error integral value from the previous cycle.

[0019] Preferably, the step of using the expected heat flux time series and the current temperature to predict the temperature change trajectory within a future time window through a server-side first-order heat conduction model, and using the initial slope of the trajectory as the differential term, includes:

[0020] Establish the first-order heat conduction differential equation for the server core:

[0021] ,in, For the core temperature, For the expected heat flux, For ambient temperature, and These are the pre-calibrated model coefficients;

[0022] The current temperature Initial values ​​of the expected heat flux time series Substituting into the equation, the initial slope is calculated. The calculated initial slope As the differential term.

[0023] Preferably, the formula for calculating the proportional gain coefficient is:

[0024] ,in, This is the proportional gain coefficient. The basic proportional gain coefficient, This represents the average expected heat flux within the future time window. This represents the absolute value of the integral of the accumulated temperature error for the current period. and This is the preset gain adjustment coefficient.

[0025] In a second aspect, a server heat dissipation control device includes the following modules:

[0026] The acquisition module is used to acquire the current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled. The task feature vector includes the instruction type composition ratio, floating-point operation density, and memory access frequency of the task. The generation module is used to predict the expected heat flux time series of the task set to be scheduled on each core based on the task feature vector and a preset instruction set heat power consumption model; calculate the thermal penalty factor positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle; generate and execute a task scheduling scheme with the goal of balancing the expected heat flux of each core and the total execution time, which includes the thermal penalty factor; the prediction module is used to calculate the current temperature of each computing core, the cumulative temperature error integral value of the previous control cycle, and the cumulative temperature error integral value of the task set to be scheduled on each core. The temperature error between the current temperature and the target temperature is calculated; an integral time step inversely proportional to the absolute value of the rate of change of temperature error is set, and the cumulative temperature error integral value of the current period is calculated using the trapezoidal integral method; using the expected heat flux time series and the current temperature, the temperature change trajectory within the future time window is predicted through the server's first-order heat conduction model, and the initial slope of the trajectory is used as the differential term; a calculation module is used to calculate the proportional gain coefficient, which is positively correlated with the absolute value of the expected heat flux and the cumulative temperature error integral value of the current period; based on the proportional gain coefficient, the cumulative temperature error integral value and the differential term, proportional, integral and differential control quantities are generated and summed to generate a heat dissipation control command for regulating the server's cooling fan speed.

[0027] In a third aspect, a computer-readable storage medium includes:

[0028] The readable storage medium stores a computer program that, when executed by a processor, implements the aforementioned server heat dissipation control method.

[0029] The beneficial effects of this invention are as follows: By inputting predictions of task thermal characteristics during the task scheduling phase and combining them with the historical thermal states of the cores, this invention generates a scheduling scheme that can balance the expected thermal load of each computing core, fundamentally avoiding the local hotspot problem caused by uneven task allocation and improving the uniformity of the overall temperature field. It utilizes a differential term constructed based on the predicted future temperature change trend from a heat conduction model to replace the traditional differential term, enhancing the predictability and anti-interference capability of the control. By inputting a proportional gain associated with the expected heat flux and calculating the cumulative error using the trapezoidal integral method, it improves the accuracy of the control response and the stability of the system, suppressing temperature overshoot and oscillation. By combining task-level thermal behavior prediction with closed-loop feedback control, it achieves more stable and coordinated regulation of server temperature. Attached Figure Description

[0030] Figure 1 The flowchart illustrating the steps of the server heat dissipation control method in this embodiment is shown in the illustration. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] like Figure 1 As shown, a server heat dissipation control method in this embodiment includes the following steps:

[0033] Step S1: Obtain the current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled. The task feature vector includes the instruction type composition ratio, floating-point operation density, and memory access frequency of the task.

[0034] Real-time temperature data is acquired by reading the digital thermal sensors (DTS) built into each computing core; the cumulative temperature error integral value stored at the end of the previous control cycle is read from memory; for each task in the set of tasks to be scheduled, offline or online analysis is performed through static code analysis or hardware performance counter (PMU) to count the number of different types of instructions such as integer operations, floating-point operations, memory accesses, and branches during execution, calculate the proportion of each type of instruction to the total number of instructions, and count the number of floating-point operations and memory accesses per unit time to obtain the task feature vector.

[0035] In an optional embodiment, the step of predicting the expected heat flux time series of the set of tasks to be scheduled executed on each core includes:

[0036] Each task in the task set is decomposed into a sequence of micro-operations within a unit time slice; a preset instruction set thermal power consumption model is queried to obtain the basic thermal power consumption value corresponding to each micro-operation; based on the proportion of instruction types, floating-point operation density, and memory access frequency in the task feature vector, the basic thermal power consumption value of each micro-operation is weighted and summed to obtain the instantaneous thermal power consumption within the time slice; the instantaneous thermal power consumption values ​​in consecutive time slices are concatenated to obtain the expected heat flux time sequence executed on the corresponding core.

[0037] Specifically, this prediction process divides a scheduled computational task into multiple unit time slices. For example, a video transcoding task is divided into multiple 10ms unit time slices based on its execution time. Within the first 10ms time slice, the execution of the task is decomposed into a series of micro-operations, such as integer addition, floating-point multiplication, and memory reads. A pre-established power consumption model database is queried, which records the basic thermal power consumption of different micro-operations, for example, integer addition is 0.1W, floating-point multiplication is 0.3W, and memory reads are 0.2W.

[0038] The feature vector of the video transcoding task is extracted, showing a high density of floating-point operations and a moderate frequency of memory accesses. Based on these features, the base thermal power consumption is adjusted; for example, the power consumption of floating-point multiplication is multiplied by a weighting factor of 1.2, resulting in 0.36W. The weighted thermal power consumption values ​​of all micro-operations within this time slice are summed to obtain a total instantaneous thermal power consumption value, such as 25W. This process is repeated for each subsequent 10ms time slice, potentially yielding values ​​such as 28W, 22W, etc. Continuing these consecutive instantaneous thermal power consumption values ​​constitutes the expected heat flux time series for the task executed on this core, such as sequence data points 25, 28, 22, etc.

[0039] In this step, not only is the current physical state (temperature) obtained, but the historical heat accumulation is also quantified by the integral value of the previous cycle. More importantly, by extracting task features (such as floating-point density), the system is given the ability to predict the future load thermal characteristics, providing a data foundation for subsequent feedforward control.

[0040] Step S2: Based on the task feature vector and the preset instruction set thermal power consumption model, predict the expected heat flux time series of the task set to be scheduled on each core; calculate the thermal penalty factor that is positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle; generate and execute the task scheduling scheme with the goal of balancing the expected heat flux of each core and the total execution time of the thermal penalty factor.

[0041] Beforehand, a series of benchmark programs are run on the target server, and power consumption and instruction execution are measured simultaneously. Using methods such as multiple linear regression, an instruction set thermal power consumption model is established, which maps task feature vectors to core power consumption. For example, thermal power consumption equals weight coefficient A multiplied by floating-point operation density, plus weight coefficient B multiplied by memory access frequency, plus a basic power consumption constant. For any task in the task set to be scheduled, the task feature vector is substituted into the model to calculate the expected thermal power consumption value when executed on a single core. This value is the expected heat flux. If the task execution time is known, a step signal with the amplitude of this heat flux value can be generated during the task duration as the expected heat flux time series.

[0042] For each computing core in the server, the cumulative temperature error integral value stored in the previous cycle is obtained, and the integral value is multiplied by a preset proportional constant to obtain the thermal penalty factor of the core. For example, if the cumulative temperature error integral value of a core in the previous cycle is large, it indicates that the core is in a high-temperature state, and the calculated thermal penalty factor is also correspondingly large.

[0043] Construct a multi-objective optimization function with two objectives: first, to minimize the expected heat flux of the largest task in the core after considering the thermal penalty factor; and second, to minimize the total completion time of the task set, i.e., the maximum completion time. The expected heat flux after considering the thermal penalty factor is equal to the sum of the expected heat fluxes of all tasks allocated to that core multiplied by one plus the thermal penalty factor of that core. Use heuristic algorithms such as genetic algorithms or particle swarm optimization to solve the optimization problem and obtain a mapping scheme for allocating tasks in the task set to specific computing cores. Call the task scheduling interface at the operating system level to bind tasks to the specified cores for execution according to this scheme.

[0044] To add a scheduling penalty to the continuously overheating core, in one optional embodiment, the thermal penalty factor is calculated as follows:

[0045] ,in, For the first One core thermal penalty factor, For the first The cumulative temperature error integral value of the previous cycle of each core. This is the preset penalty gain coefficient.

[0046] For example, suppose there are two cores, core 1 and core 2, and a preset penalty gain coefficient. The value is 0.05. After a control cycle ends, check the cumulative temperature error integral value of each core from the previous cycle. If core 1 has been operating under high load for a long time, the cumulative temperature error integral value from the previous cycle will be 0.05. A value of -50 degrees Celsius per second indicates that the temperature remains above the target value.

[0047] Calculate the thermal penalty factor of core 1 according to the formula. The thermal penalty factor of core 1 is obtained. The value is 3.5. For core 2, if the integral value of the previous cycle is +10 degrees per second, it indicates that the temperature is lower than the target value, and the penalty factor for core 2 is calculated accordingly. A penalty factor of 1.5 will reduce the priority of new tasks being assigned to core 1 in subsequent task scheduling decisions.

[0048] In an optional embodiment, a task scheduling scheme is generated and executed with the goal of balancing the expected heat flux of each core factor that includes the thermal penalty factor with the total execution time, including:

[0049] Construct the optimization objective function ,in, The total completion time for the task set. For the first One core thermal penalty factor, To be allocated to the first The expected total heat production of all tasks in each core, and For preset weighting coefficients, and These are the normalized time and normalized heat, respectively, used to make the components of the objective function dimensionless;

[0050] The particle swarm optimization algorithm is used to iteratively optimize the allocation scheme of tasks on each core until the objective function is achieved. The value of converges to the minimum, thus obtaining the task scheduling scheme.

[0051] Specifically, assuming three tasks need to be scheduled to two cores, the weighting coefficients are... and The values ​​are set to 0.6 and 0.4 respectively. Based on the thermal penalty factors calculated above (thermal penalty factor of core 1),... It is 3.5, with 2 cores. (1) The particle swarm optimization algorithm randomly generates multiple scheduling schemes as initial particles. For example, scheme A assigns tasks 1 and 2 to core 1, and task 3 to core 2. The total completion time is calculated based on this scheme. The total heat generated by core 1 is allocated within 100ms. For 500 Total heat production of core 2 For 200 .

[0052] Calculate the objective function of scheme A Value. Assuming normalized time. 120ms, normalized heat For 1000 The objective function can be calculated. The value is 1.28. The algorithm will continue to evaluate other options, such as option B, which assigns task 1 to core 1 and tasks 2 and 3 to core 2, and calculate... Value. Through multiple iterations, the particle swarm optimization algorithm continuously adjusts task allocation to find the optimal value. The option with the smallest value is selected and adopted as the implementation option because it represents the best balance between speed and thermal management.

[0053] In this step, by predicting heat flux and introducing a thermal penalty factor, the scheduling algorithm no longer focuses solely on performance but actively avoids "hot" cores, distributing the thermal load evenly across all computing units. This effectively prevents localized hotspots caused by task backlog, reduces the pressure on subsequent cooling systems, and improves chip reliability.

[0054] Step S3: Calculate the temperature error between the current temperature and the target temperature; set an integral time step that is inversely proportional to the absolute value of the temperature error change rate, and calculate the cumulative temperature error integral value of the current period using the trapezoidal integral method; use the expected heat flux time series and the current temperature to predict the temperature change trajectory within the future time window using the server's first-order heat conduction model, and use the initial slope of the trajectory as the differential term.

[0055] The steps for calculating the cumulative temperature error integral value for the current period include:

[0056] Calculate the current temperature error Temperature error compared to the previous cycle The difference is used to obtain the change in temperature error. ;

[0057] According to the formula Calculate the integration time step ,in, Based on the time step, This is the preset integral step size adjustment coefficient;

[0058] Through formula Calculate the cumulative temperature error integral value for the current period. ,in, This is the cumulative temperature error integral value from the previous cycle.

[0059] For example, assume a base time step The integral step size adjustment coefficient is 10ms. It is 0.5. Period, measuring temperature error It is 2℃. At present... During the cycle, due to the start of a high-load task, the temperature rose rapidly, resulting in a measured temperature error. The temperature is 8℃. At this point, the change in temperature error... The temperature is 6℃. Due to the drastic temperature changes, a smaller integration step size is needed to improve accuracy.

[0060] The integration time step can be calculated using the formula. The integration time step is 2.5ms, which is relatively short. The cumulative temperature error integral value is updated using the trapezoidal integration method. If the cumulative value from the previous cycle... If the time interval is 20 degrees per millisecond, then the cumulative value for the current period is... It is 32.5 degrees per millisecond. If the temperature change is gradual, for example... The calculated value is 0.2℃. This will be close to 10ms, thus reducing unnecessary calculations while maintaining accuracy.

[0061] Using the expected heat flux time series and the current temperature, a server-side first-order heat conduction model is used to predict the temperature change trajectory within a future time window, with the initial slope of the trajectory as the differential term, including:

[0062] Establish the first-order heat conduction differential equation for the server core:

[0063] ,in, For the core temperature, For the expected heat flux, For ambient temperature, and These are the pre-calibrated model coefficients;

[0064] The current temperature Initial values ​​of the expected heat flux time series Substituting into the equation, the initial slope is calculated. ;

[0065] The calculated initial slope As the differential term.

[0066] Specifically, a predicted value about future temperature trends is provided to the PID controller, assuming that the thermal model coefficients of a server core have been pre-calibrated, and the heat generation coefficient... The heat dissipation coefficient is 0.15. The value is 0.03. Current ambient temperature. The core's current temperature is 25°C, as measured by a sensor. The temperature is 75℃. Based on the upcoming scheduled task, the expected heat flux time series for a period of time is predicted, with the first value being... It is 40W.

[0067] The initial slope is calculated by substituting the real-time and predicted data into the first-order heat conduction differential equation. .get The value is 4.5. The unit of the obtained value 4.5 is °C / s, which represents the rate of change of the core temperature under the current state and expected load. The differential term... This input is fed into the derivative of the PID controller, enabling the controller to respond to the expected temperature rise trend in advance, rather than waiting for the temperature error to actually occur before making adjustments.

[0068] In this step, by adopting an integration step size that is inversely proportional to the rate of change of error, details can be captured when the temperature fluctuates drastically, while integration saturation can be avoided in steady state. The differential term is calculated using a first-order heat conduction model, and the trend is predicted by using physical laws rather than simply data differences, which greatly improves the ability to predict future temperature changes and enhances the ability to resist noise interference.

[0069] Step S4: Calculate the proportional gain coefficient that is positively correlated with the absolute value of the expected heat flux and the cumulative temperature error integral value of the current cycle; generate proportional, integral and differential control quantities based on the proportional gain coefficient, the cumulative temperature error integral value and the differential term, and sum them to generate a heat dissipation control command for adjusting the speed of the server cooling fan.

[0070] Specifically, a base proportional gain coefficient is set, plus a positive coefficient multiplied by the expected heat flux of the task to be executed on the current core, plus another positive coefficient multiplied by the absolute value of the integral value of the cumulative temperature error in the current cycle, and the sum of the three is used as the proportional gain coefficient; when a high-power task is about to be executed or the historical cumulative temperature difference is large, the proportional gain coefficient will increase accordingly, making the control effect stronger.

[0071] Multiply the calculated proportional gain coefficient by the current temperature error to obtain the proportional term; multiply a fixed preset integral gain coefficient by the cumulative temperature error integral value of the current period to obtain the integral term; multiply a fixed preset differential gain coefficient by the differential term to obtain the differential term. Summate the calculated proportional, integral, and differential control values ​​algebraically to obtain a comprehensive control output value; convert the comprehensive control output value into a pulse width modulation (PWM) duty cycle signal for the corresponding fan using a preset mapping function or lookup table method, for example, an integer from 0 to 255; send this PWM signal to the fan controller via the server's baseboard management controller (BMC) to adjust the actual fan speed.

[0072] In an optional embodiment, the formula for calculating the proportional gain coefficient is:

[0073] ,in, This is the proportional gain coefficient. The basic proportional gain coefficient, This represents the average expected heat flux within the future time window. This represents the absolute value of the integral of the accumulated temperature error for the current period. and This is the preset gain adjustment coefficient.

[0074] Specifically, adjust the proportional gain in the PID controller. It can adapt to different workloads and system thermal states. Set the base proportional gain coefficient. The gain adjustment coefficient is 0.8. It is 0.01. The value is 0.002. Within one control period, the average heat flux for a future time window is predicted. A value of 50W indicates an upcoming high-load scenario. Calculate the cumulative temperature error integral value for the current cycle. -60 degrees per second, absolute value The value of 60 indicates that the core has been deviating from the target temperature for some time.

[0075] Calculate the proportional gain coefficient according to the formula. Larger This value allows the PID controller to respond more strongly and quickly to the current temperature error. An aggressive control strategy is necessary under high loads and existing temperature deviations to prevent temperature runaway. Conversely, under low loads and stable temperatures, the calculated... The value will be smaller, making the control effect more gentle.

[0076] This step achieves dynamic matching of control intensity. The proportional gain coefficient is positively correlated with the predicted heat flux and cumulative error, meaning that when facing an upcoming high load or significant historical heat buildup, the system will output a more aggressive control input to quickly suppress temperature rise; while maintaining gentle control under low loads. This non-linear adjustment mechanism optimizes fan efficiency and reduces unnecessary noise and vibration while ensuring that the temperature does not exceed the limit.

[0077] In a second embodiment, the present invention also provides a server heat dissipation control device, comprising the following modules:

[0078] The acquisition module is used to acquire the current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled. The task feature vector includes the instruction type composition ratio, floating-point operation density and memory access frequency of the task.

[0079] The generation module is used to predict the expected heat flux time series of the task set to be scheduled on each core based on the task feature vector and the preset instruction set thermal power consumption model; calculate the thermal penalty factor that is positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle; and generate and execute the task scheduling scheme with the goal of balancing the expected heat flux of each core and the total execution time, which includes the thermal penalty factor.

[0080] The prediction module is used to calculate the temperature error between the current temperature and the target temperature; it sets an integral time step that is inversely proportional to the absolute value of the temperature error change rate, and uses the trapezoidal integral method to calculate the cumulative temperature error integral value of the current period; using the expected heat flux time series and the current temperature, it predicts the temperature change trajectory within the future time window through the server's first-order heat conduction model, and uses the initial slope of the trajectory as the differential term.

[0081] The calculation module is used to calculate a proportional gain coefficient that is positively correlated with the absolute value of the expected heat flux and the cumulative temperature error integral value of the current period; based on the proportional gain coefficient, the cumulative temperature error integral value and the differential term, it generates proportional, integral and differential control quantities and sums them to generate a heat dissipation control command for adjusting the speed of the server cooling fan.

[0082] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the server heat dissipation control method described above.

[0083] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0084] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0085] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for regulating server heat dissipation, characterized in that, Includes the following steps: Obtain the current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled. The task feature vector includes the instruction type composition ratio, floating-point operation density and memory access frequency of the task. Based on the task feature vector and the preset instruction set thermal power consumption model, predict the expected heat flux time series of the task set to be scheduled on each core; calculate the thermal penalty factor that is positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle. With the objective of balancing the expected heat flux of each core component and the total execution time by incorporating the thermal penalty factor, a task scheduling scheme is generated and executed; the thermal penalty factor satisfies: , For the first One core thermal penalty factor, For the first The cumulative temperature error integral value of the previous cycle of each core. This is the preset penalty gain coefficient; Calculate the temperature error between the current temperature and the target temperature; Set an integration time step inversely proportional to the absolute value of the rate of change of temperature error, and calculate the cumulative temperature error integral value for the current period using the trapezoidal integration method, including: Calculate the current temperature error Temperature error compared to the previous cycle The difference is used to obtain the change in temperature error. ; According to the formula Calculate the integration time step , Based on the time step, This is the preset integral step size adjustment coefficient; Through formula Calculate the cumulative temperature error integral value for the current period. , This is the integral value of the cumulative temperature error from the previous cycle; Using the expected heat flux time series and the current temperature, the temperature change trajectory within the future time window is predicted through a first-order heat conduction model on the server, and the initial slope of the trajectory is used as the differential term. Calculate the proportional gain coefficient, which is positively correlated with the absolute value of the cumulative temperature error integral value of the current cycle. Based on the proportional gain coefficient, the cumulative temperature error integral value, and the differential term, generate proportional, integral, and differential control quantities and sum them to generate a heat dissipation control command for adjusting the speed of the server cooling fan.

2. The server heat dissipation control method according to claim 1, characterized in that, The step of predicting the expected heat flux time series of the scheduled task set executed on each core includes: Each task in the task set is broken down into a sequence of micro-operations within a unit time slice; Query the preset instruction set thermal power consumption model to obtain the basic thermal power consumption value corresponding to each micro-operation; Based on the proportion of instruction types, floating-point operation density, and memory access frequency in the task feature vector, the basic thermal power consumption values ​​of each micro-operation are weighted and summed to obtain the instantaneous thermal power consumption within the time slice. By concatenating the instantaneous thermal power values ​​within consecutive time slices, the expected thermal flux time series executed on the corresponding core is obtained.

3. The server heat dissipation control method according to claim 1, characterized in that, The process of generating and executing a task scheduling scheme with the goal of balancing the expected heat flux of each core component and the total execution time, which are included in the thermal penalty factor, includes: Construct the optimization objective function ,in, The total completion time for the task set. For the first One core thermal penalty factor, To be allocated to the first The expected total heat production of all tasks in each core, and For preset weighting coefficients, and These are the normalized time and normalized heat, respectively, used to make the components of the objective function dimensionless; The particle swarm optimization algorithm is used to iteratively optimize the allocation scheme of tasks on each core until the objective function is achieved. The value of converges to the minimum, thus obtaining the task scheduling scheme.

4. The server heat dissipation control method according to claim 1, characterized in that, The step of using the expected heat flux time series and the current temperature to predict the temperature change trajectory within a future time window through a server-side first-order heat conduction model, and using the initial slope of the trajectory as the differential term, includes: Establish the first-order heat conduction differential equation for the server core: ,in, for The core temperature of every moment for Expected heat flux at any given time. For ambient temperature, and These are the pre-calibrated model coefficients; The current temperature Initial values ​​of the expected heat flux time series Substituting into the equation, the initial slope is calculated. ; The calculated initial slope As the differential term.

5. The server heat dissipation control method according to claim 1, characterized in that, The formula for calculating the proportional gain coefficient is as follows: ,in, This is the proportional gain coefficient. The basic proportional gain coefficient, This represents the average expected heat flux within the future time window. This represents the absolute value of the integral of the accumulated temperature error for the current period. and This is the preset gain adjustment coefficient.

6. A server heat dissipation control device for implementing the server heat dissipation control method according to any one of claims 1 to 5, characterized in that, Includes the following modules: The acquisition module is used to acquire the current temperature of each computing core of the server, the cumulative temperature error integral value of the previous control cycle, and the task feature vector of the task set to be scheduled. The task feature vector includes the instruction type composition ratio, floating-point operation density and memory access frequency of the task. The generation module is used to predict the expected heat flux time series of the task set to be scheduled on each core based on the task feature vector and the preset instruction set thermal power consumption model. Calculate the thermal penalty factor that is positively correlated with the cumulative temperature error integral value of each computing core in the previous control cycle; With the goal of balancing the expected heat flux of each core component and the total execution time by taking into account the thermal penalty factor, a task scheduling scheme is generated and executed. The prediction module is used to calculate the temperature error between the current temperature and the target temperature; it sets an integral time step that is inversely proportional to the absolute value of the temperature error change rate, and uses the trapezoidal integral method to calculate the cumulative temperature error integral value of the current period; using the expected heat flux time series and the current temperature, it predicts the temperature change trajectory within the future time window through the server's first-order heat conduction model, and uses the initial slope of the trajectory as the differential term. The calculation module is used to calculate a proportional gain coefficient that is positively correlated with the absolute value of the expected heat flux and the cumulative temperature error integral value of the current period; based on the proportional gain coefficient, the cumulative temperature error integral value and the differential term, it generates proportional, integral and differential control quantities and sums them to generate a heat dissipation control command for adjusting the speed of the server cooling fan.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the server heat dissipation control method as described in any one of claims 1-5.

Images