Control system of efficient and energy-saving machine room of data center

Abstract

The invention discloses a control system for an efficient energy-saving machine room of a data center, and belongs to the technical field of machine room control systems, and the control system specifically comprises the steps: collecting the air flow track of the machine room and the surface temperature distribution of equipment, and forming an airflow thermal field state set containing a velocity field and a temperature field; the calculation instruction type, the data access mode and the floating point calculation intensity of the server are collected to form a calculation feature set. And carrying out space-time alignment on an airflow track and temperature distribution, establishing an unsteady airflow-thermal field coupling equation set, converting floating point operation intensity into heat source item input, and carrying out iterative solution to obtain a future three-dimensional thermal field evolution sequence. Analyzing the sequence temperature gradient, identifying a heat retention area which is continuously lower than a threshold value, generating a guide plate angle control parameter and a calculation task migration parameter according to the heat retention area, converting the parameters into a driving signal and a task redistribution instruction, and transmitting the driving signal and the task redistribution instruction to execution equipment; the heat dissipation efficiency and the equipment stability are improved by accurately regulating and controlling the equipment parameters of the machine room.

Description

Technical Field

[0001] This invention relates to the field of data center control system technology, and specifically to a control system for a high-efficiency and energy-saving data center. Background Technology

[0002] In existing technologies, temperature management in data center server rooms primarily employs temperature sensor-based feedback control systems. These systems monitor ambient temperature in real time using temperature sensors deployed throughout the server room. When the temperature in a certain area exceeds a set threshold, the operating parameters of the corresponding cooling equipment in that area are automatically adjusted. Some improved solutions also incorporate computing equipment load monitoring, dynamically adjusting cooling intensity based on server utilization to create a load-aware cooling strategy. These solutions, to some extent, achieve on-demand allocation of cooling resources and improve energy efficiency.

[0003] However, existing methods primarily rely on fixed duct designs and temperature threshold-based cooling control, which cannot adapt to dynamically changing airflow patterns. Traditional monitoring methods are limited to temperature sensor readings, lacking precise capture of airflow trajectories, resulting in blind spots in the identification of heat stagnation areas. Furthermore, current technologies fail to establish a dynamic correlation between computational task characteristics and thermal field evolution, leading to a disconnect between cooling adjustments and changes in computational load. Due to the lack of quantitative analysis of the unsteady airflow-thermal field coupling relationship, existing systems struggle to achieve precise airflow guidance control, resulting in low cooling efficiency and energy waste. Summary of the Invention

[0004] The purpose of this invention is to provide a control system for a high-efficiency and energy-saving data center, solving the following technical problems: Existing data center cooling control methods cannot adapt to dynamically changing airflow organization, and temperature threshold-based control methods suffer from response lag. At the same time, the lack of quantitative analysis on the dynamic correlation between computing task characteristics and thermal field evolution leads to a disconnect between cooling regulation and changes in computing load, resulting in energy waste.

[0005] The objective of this invention can be achieved through the following technical solutions: A control system for a high-efficiency and energy-saving data center includes: The airflow state acquisition module is used to collect airflow trajectory data and equipment surface temperature distribution data in the computer room space, forming an airflow thermal field state set that includes velocity field and temperature field; The computational feature acquisition module is used to collect computational instruction type data, data access mode data, and floating-point operation intensity data during server operation, forming a set of computational features that reflect the characteristics of computational load. The thermal field evolution prediction module is used to align airflow trajectory data with equipment surface temperature distribution data in time and space, establish an unsteady airflow-thermal field coupling equation set, and convert floating-point operation intensity data into heat source terms to input the equation set. Through iterative solution, a three-dimensional thermal field evolution sequence for future time periods is obtained. The thermal retention analysis module is used to analyze the temperature gradient distribution in the three-dimensional thermal field evolution sequence, identify areas where the temperature gradient is continuously below a set threshold, and mark them as thermal retention areas. The flow control module is used to generate flow deflector angle control parameters and calculation task migration parameters based on the spatial distribution characteristics of the heat retention area. The instruction execution module is used to convert the guide vane angle control parameters into guide vane drive signals, convert the calculation task migration parameters into task reassignment instructions, and transmit them to the execution device.

[0006] As a further aspect of the present invention: the specific process of collecting airflow trajectory data and equipment surface temperature distribution data in the airflow state acquisition module is as follows: Multiple measurement points are set up in the computer room, the coordinate position of each measurement point is recorded, the air flow velocity and temperature value of each measurement point are collected, the air flow velocity is converted into velocity vector data, and the temperature value is mapped to the coordinate position of the corresponding measurement point. The velocity vector data and temperature values ​​are organized in chronological order. Data with inconsistent timestamps are time-aligned. Missing data is interpolated using values ​​from adjacent time points. The processed velocity vector data and temperature values ​​are stored as a three-dimensional array containing time, space, and data dimensions.

[0007] As a further aspect of the present invention: the specific process of acquiring data during server operation in the computational feature acquisition module is as follows: The system collects data by statistically analyzing the number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions to form computational instruction type data; records cache access frequency, memory read / write frequency, and storage input / output operation frequency to form data access pattern data; calculates the ratio of floating-point operations to data accesses per unit time to form floating-point arithmetic intensity data; categorizes the collected data according to processor cores, adds a collection time stamp to each data item, and uses a hierarchical data table to store the computational feature set. The first level of the hierarchical data table identifies the processor number, and the second level stores the time series of various feature data.

[0008] As a further aspect of the present invention: the specific process of obtaining the three-dimensional thermal field evolution sequence in the thermal field evolution prediction module is as follows: Airflow trajectory data is resampled to a non-uniform computing grid, wherein the grid spacing in densely populated areas is less than a set threshold, and the grid spacing in sparsely populated areas is greater than a set threshold; a heat flux boundary condition based on the measured temperature distribution is applied at the boundary of the computing domain. An unsteady coupled equation system incorporating buoyancy and turbulence effects is constructed. A dynamic heat source term derived from floating-point arithmetic intensity data is introduced into the unsteady coupled equation system. A step-by-step solution strategy is adopted to first solve the flow field equations, and then input the flow field solution results into the temperature field equations. Multiple sub-iteration calculations are performed within each time step to output three-dimensional temperature field data, three-dimensional velocity field data, and three-dimensional pressure field data at multiple future time points.

[0009] As a further aspect of the present invention, the specific process of deriving the dynamic heat source term is as follows: Establish the correspondence between computing instruction type and heat generation per unit time, determine the heat-to-work conversion coefficient of floating-point arithmetic unit based on processor architecture parameters, collect processor operating frequency and voltage data, and calculate instantaneous heat generation power in combination with the heat-to-work conversion coefficient; Based on the spatial coordinates of the processor, the instantaneous heat generation power is distributed to the corresponding grid cells. The thermal resistance parameters of the heat sink and the thermal conductivity of the interface material are read, and the distribution characteristics of the heat source in the grid cells are calculated to form heat source distribution data containing spatial location attributes and time variation characteristics. The heat source distribution data is used as a dynamic heat source term input into the unsteady coupled equation system.

[0010] As a further aspect of the present invention: the specific process for identifying the thermal retention region in the thermal retention analysis module is as follows: Temperature field data for multiple consecutive time steps are extracted from the three-dimensional thermal field evolution sequence. The temperature change rate of each grid point in the time dimension is calculated, and the temperature gradient value between each grid point and its neighboring grid points is calculated. Grid points with a temperature change rate lower than the change rate threshold and a temperature gradient value lower than the gradient threshold are marked as candidate points. The candidate points are spatially clustered to form continuous regions. The volume, average temperature and maximum temperature of each continuous region are calculated. Continuous regions with a volume greater than the volume threshold and an average temperature higher than the temperature threshold are marked as thermal stagnation regions. The spatial range data of the thermal stagnation regions are output.

[0011] As a further aspect of the present invention: the specific process of generating the guide vane angle control parameters and calculating the task migration parameters in the guide vane control module is as follows: The spatial distribution characteristics of the heat retention area are analyzed, including the region's geometry, spatial location, and volume data. Based on the characteristics of the heat retention area, the deflector control mapping table is queried to obtain the initial angle setting. The initial angle setting is then corrected by combining real-time airflow velocity data and temperature gradient data to generate the deflector angle control parameters. Identify the computing tasks running on computing nodes within the thermal stagnation area and obtain the computing characteristic data of the tasks; determine the task migration priority based on the floating-point operation intensity in the computing characteristic data, select the target computing node whose temperature distribution meets the requirements and whose resource utilization is below the threshold, and generate computing task migration parameters containing task identifier, source node address and target node address.

[0012] As a further aspect of the present invention: the process of constructing the guide vane control mapping table is as follows: A parametric mesh is established based on the 3D model of the computer room. The mesh cell size is consistent with the identification accuracy of the heat retention area. The flow field distribution and temperature field distribution under different combinations of guide vane angles are obtained through computational fluid dynamics simulation. A database of the correspondence between guide vane angle and heat retention area elimination effect is established. A supervised learning method is used to train the angle recommendation model. The training features include the location, volume, and temperature gradient of the thermal stagnation region. During actual operation, the control effect data of the deflector is collected, including temperature changes and airflow organization changes before and after angle setting. An online update mechanism based on reinforcement learning is established. The angle recommendation strategy is optimized according to the control effect data. The performance of the updated deflector control mapping table under actual operating conditions is periodically verified.

[0013] As a further aspect of the present invention: the specific process of transmitting instructions to the execution device in the instruction execution module is as follows: The guide vane angle parameters are encoded into a stepper motor drive protocol that includes angle calibration instructions and error compensation instructions; the computation task migration parameters are encapsulated into a distributed system communication protocol that includes task state saving instructions and task state recovery instructions. By employing a priority scheduling mechanism to transmit the flow guide plate control signal in real-time via industrial Ethernet, and by using a data center network with redundant links to transmit the task migration instructions, a monitoring loop for the instruction execution status that collects flow guide plate position feedback and task migration progress in real time is established. A retransmission mechanism is designed to set an upper limit for the number of retransmissions based on the instruction type.

[0014] The beneficial effects of this invention are: This invention establishes an unsteady airflow-thermal field coupling equation set, converting floating-point operation intensity data into dynamic heat source terms, thereby achieving accurate prediction of the three-dimensional thermal field evolution sequence and effectively overcoming the hysteresis defect of traditional temperature feedback control. By analyzing the temperature gradient distribution and airflow velocity field, an isothermal surface extraction algorithm is used to identify thermal stagnation regions, solving the problem of inaccurate identification of thermal stagnation regions by traditional monitoring methods. Based on the deflector control mapping table, angle control parameters are generated, and combined with computational task migration to achieve collaborative control, establishing a dynamic correlation between computational task characteristics and thermal field evolution, eliminating the response disconnect between cooling adjustment and computational load changes. The use of non-uniform grid technology and step-by-step solution strategy improves the accuracy of thermal field prediction, and precise airflow guidance is achieved through deflector angle optimization and task migration mechanism, significantly improving cooling efficiency and reducing energy consumption while preventing local overheating. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the present invention is a control system for a high-efficiency and energy-saving data center, comprising: The airflow state acquisition module is used to collect airflow trajectory data and equipment surface temperature distribution data in the computer room space, forming an airflow thermal field state set that includes velocity field and temperature field; The computational feature acquisition module is used to collect computational instruction type data, data access mode data, and floating-point operation intensity data during server operation, forming a set of computational features that reflect the characteristics of computational load. The thermal field evolution prediction module is used to align airflow trajectory data with equipment surface temperature distribution data in time and space, establish an unsteady airflow-thermal field coupling equation set, and convert floating-point operation intensity data into heat source terms to input the equation set. Through iterative solution, a three-dimensional thermal field evolution sequence for future time periods is obtained. The thermal retention analysis module is used to analyze the temperature gradient distribution in the three-dimensional thermal field evolution sequence, identify areas where the temperature gradient is continuously below a set threshold, and mark them as thermal retention areas. The flow control module is used to generate flow deflector angle control parameters and calculation task migration parameters based on the spatial distribution characteristics of the heat retention area. The instruction execution module is used to convert the guide vane angle control parameters into guide vane drive signals, convert the calculation task migration parameters into task reassignment instructions, and transmit them to the execution device.

[0019] In a preferred embodiment of the present invention, the specific process of collecting airflow trajectory data and equipment surface temperature distribution data in the airflow state acquisition module is as follows: First, the layout of the measurement points is planned. Based on the actual layout of the computer room, a network of measurement points is established at key locations. These points cover the areas of the server rack air inlets, exhaust outlets, air conditioning supply vents, and air conditioning return vents. Each measurement point has a unique coordinate identifier, represented by a three-dimensional coordinate system. The origin of the coordinate system is set at a corner of the computer room, with the X-axis extending along the length of the room, the Y-axis along the width, and the Z-axis along the height. The spacing between the measurement points is determined based on the required monitoring accuracy, with increased density in areas with drastic airflow changes.

[0020] After the measurement points are set up, the corresponding sensing equipment is installed. Each measurement point is equipped with a wind speed sensor and a temperature sensor. The wind speed sensor uses a hot-wire anemometer, which can measure the speed and direction of airflow. The temperature sensor uses a platinum resistance thermometer, which has high measurement accuracy and stability. All sensors are connected to the data acquisition unit via a fieldbus, and the data acquisition unit is responsible for converting the sensor signals into digital quantities.

[0021] The data acquisition process follows a pre-defined sampling period. At the start of each sampling period, the data acquisition unit sends acquisition commands to all sensors. The wind speed sensor measures the airflow velocity at the current point, including the velocity value and direction angle. The temperature sensor measures the temperature value at the current point. The acquired raw data includes sensor number, measurement timestamp, and measurement value, among other information.

[0022] After obtaining the raw data, data preprocessing is required. For airflow velocity data, the velocity values ​​and direction angles are converted into velocity vector representations. The velocity vector contains three components, representing the velocity components along the X, Y, and Z axes. For temperature data, a mapping relationship is established between temperature values ​​and the coordinate positions of the measurement points, forming temperature distribution data with spatial location information.

[0023] Next, time series processing is performed. Data collected from different measurement points are arranged in chronological order. Since there may be slight time differences in the data collected by different sensors, time alignment is required for data with inconsistent timestamps. Time alignment uses a linear interpolation method, with the master clock's time point as the reference, to calculate the value of each sensor's data at the reference time point.

[0024] During data processing, missing data for some measurement points may be discovered. To handle missing data, interpolation is performed using values ​​from adjacent time points. Specifically, if data for a measurement point is missing at a certain time point, the arithmetic mean of the measurement values ​​from the preceding and following time points is calculated as the supplementary value for that time point.

[0025] After data preprocessing, the processed velocity vector data and temperature values ​​are stored as a three-dimensional array structure. The first dimension of the three-dimensional array represents the time series, arranged in chronological order of sampling time. The second dimension represents the spatial location, indexed by the coordinates of the measurement points. The third dimension represents the data type, including the three components of the velocity vector and the temperature value. This storage structure facilitates subsequent data access and processing.

[0026] Data storage employs a circular buffer approach, where new data continuously overwrites older data to maintain data freshness. Simultaneously, important data is periodically transferred to permanent storage for historical data analysis and system optimization.

[0027] In another preferred embodiment of the present invention, the specific process of collecting data from the server during runtime in the computational feature acquisition module is as follows: First, a monitoring unit is deployed in the processor instruction pipeline. This monitoring unit, implemented using hardware performance counters, can count the number of instructions executed in real time. Specifically, the monitored instruction types include integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions. Integer arithmetic instructions include arithmetic-logical arithmetic instructions and shift instructions. Floating-point arithmetic instructions include floating-point addition instructions and floating-point multiplication instructions. Vector arithmetic instructions include single-instruction multiple-data instructions. The performance counter updates the instruction count every clock cycle and outputs the count result to the acquisition buffer at fixed time intervals.

[0028] Simultaneously, an access monitoring unit is deployed in the memory controller. This unit records cache access events, memory read / write events, and storage I / O operation events. Cache access events track the number of cache line hits and misses. Memory read / write events track the frequency of memory page accesses. Storage I / O operation events track the number of read / write operations performed on the solid-state drive (SSD) or hard disk. These statistics reflect the data access characteristics of the computing task.

[0029] To obtain floating-point operation intensity data, it is necessary to monitor both the activity of the floating-point unit and the data bus transmission status. The floating-point unit activity is obtained through a performance counter, recording the number of floating-point operations performed per unit time. The data bus transmission status is obtained through a bus monitor, recording the amount of data transmitted per unit time. Floating-point operation intensity is defined as the ratio of the number of floating-point operations to the amount of data accessed; this ratio reflects the computational intensity of the task.

[0030] The collected data needs to be categorized and processed according to processor cores. Each processor core has a unique identifier, and the corresponding processor core identifier is recorded during data collection. For multi-core processors, monitoring data for each core is collected and stored independently. During data categorization, data from the same processor core are grouped together to form an independent dataset for each core.

[0031] Each data item requires an accompanying acquisition timestamp. The timestamp uses a high-precision clock to ensure a consistent time reference for data from different processor cores. The timestamp accuracy must be at the microsecond level to meet the time synchronization requirements of subsequent data analysis.

[0032] Finally, a hierarchical data table is used to store the computational feature set. The first level of the hierarchical data table is the processor number index, with each processor number corresponding to a sub-table. The second level is the time series of the feature data, storing various types of feature data in chronological order. Specifically, the stored feature data includes computation instruction type data, data access pattern data, and floating-point operation intensity data. Computation instruction type data includes the number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions. Data access pattern data includes cache access frequency, memory read / write frequency, and storage input / output operation frequency. Floating-point operation intensity data includes the ratio of the number of floating-point operations to the number of data accesses per unit time.

[0033] The data table uses a columnar storage structure, grouping data of the same type together for easy and rapid querying and analysis. The table also includes metadata, recording parameters such as the start and end times of data collection and the sampling frequency. Updates are performed append-only, with new data records continuously added to the end of the table to maintain data integrity and continuity.

[0034] To ensure data reliability, the system also incorporates a data verification mechanism. After each data update, a checksum is calculated to prevent errors during transmission or storage. Furthermore, the system periodically performs integrity checks on stored data to promptly identify and correct any errors.

[0035] In another preferred embodiment of the present invention, the specific process of obtaining the three-dimensional thermal field evolution sequence in the thermal field evolution prediction module is as follows: First, a non-uniform computational grid is established. Based on the equipment distribution within the server room, the space is divided into multiple grid cells. In densely populated server rack areas, a smaller grid spacing, such as 0.1 meters, is used. In aisles and unused areas, a larger grid spacing, such as 0.5 meters, is used. Hexahedral elements are used for mesh generation, with each cell having 8 nodes. During mesh generation, it is crucial to ensure a smooth transition in grid size between adjacent areas, avoiding abrupt size changes. After mesh generation, each grid cell is assigned a unique number, and its spatial coordinate range is recorded.

[0036] Next, data resampling is performed. The collected airflow trajectory data is mapped onto grid nodes. For each grid node, measurement points within a certain radius are located, for example, all measurement points within a 1-meter radius. The velocity vector value of the node is calculated using an inverse distance weighting method, with closer measurement points having higher weights. For temperature data, the same method is used to interpolate the temperature values ​​of the measurement points onto the grid nodes. After resampling, each grid node contains both a velocity vector and a temperature value.

[0037] Set heat flux boundary conditions at the boundaries of the computational domain. These boundaries include the walls, ceiling, and floor of the computer room. Specify a temperature value or heat flux density for each boundary element based on measured temperature data. For air conditioning supply vents, set a velocity inlet boundary condition, specifying the supply air velocity and temperature. For return air vents, set a pressure outlet boundary condition. Other solid walls use adiabatic or given heat flux density boundary conditions.

[0038] When constructing an unsteady coupled equation system, buoyancy and turbulence effects need to be considered. The governing equations include the mass conservation equation, momentum conservation equation, and energy conservation equation. A buoyancy source term is added to the momentum equation; the magnitude of buoyancy is proportional to the temperature difference. Turbulence effects are considered by introducing a turbulence model, such as the standard k-ε turbulence model. The physical properties in the equation system include air density, specific heat capacity, thermal conductivity, and viscosity, which vary with temperature.

[0039] The derivation process of the dynamic heat source term is as follows. First, the correspondence between the type of calculation instruction and the heat generation per unit time is established. Through experimental testing, the heat generation characteristics of different types of instructions are determined. For example, 1 million integer arithmetic instructions generate 0.1 watts of heat per 1 million executions, floating-point arithmetic instructions generate 0.5 watts of heat per 1 million executions, and vector arithmetic instructions generate 1 watt of heat per 1 million executions. These correspondences are stored in a lookup table.

[0040] The thermal-to-power conversion factor (TDC) is determined based on the processor's architecture parameters. The TDC is related to its manufacturing process and core area. For example, a processor using a 7nm process has a TDC of 0.9. Frequency and voltage data during processor operation are collected; frequency is in GHz and voltage is in volts. Instantaneous heat generation is calculated using the TDC and the frequency multiplied by the square of the voltage, then multiplied by the TDC.

[0041] Based on the processor's spatial coordinates, instantaneous heat generation power is distributed to the corresponding grid cells. First, the grid cell number where the processor is located is determined, and then the heat generation power is evenly distributed to all nodes in that cell. If the processor spans multiple grid cells, the power is distributed proportionally by volume.

[0042] Read the thermal resistance parameters of the heat sink and the thermal conductivity of the interface material. The unit of heat sink thermal resistance is °C / W, and the unit of interface material thermal conductivity is W / m·℃. Based on these parameters, calculate the distribution characteristics of the heat source within the grid cells. Considering the heat conduction process, the generated power is first conducted to the heat sink through the interface material, and then dissipated into the air through convection. This process affects the spatial distribution of the heat source.

[0043] When generating heat source distribution data, time-varying characteristics need to be considered. The heat source distribution data includes the heat source intensity value for each grid cell at each time step. The heat source intensity value is calculated based on the processor's real-time power, taking into account the effect of thermal inertia. Thermal inertia causes a certain delay in heat source changes relative to power changes; this delay is simulated using a first-order inertial element.

[0044] Heat source distribution data is input as dynamic heat source terms into the unsteady coupled equations. Heat source intensity values ​​are added to the source terms of the energy equations. Each grid cell has a corresponding heat source term, which is continuously updated as the time step progresses.

[0045] A step-by-step solution strategy is employed to solve the coupled equations. First, the flow field equations, including the mass conservation equation and the momentum conservation equation, are solved. A semi-implicit method for the pressure coupling equations is then used to solve for the velocity and pressure fields. Within each time step, the momentum equation is solved first to obtain the intermediate velocity field, then the pressure correction equation is solved, and finally, the velocity and pressure fields are updated.

[0046] The flow field solution is input into the temperature field equation. The velocity and pressure fields are substituted as known quantities into the energy equation. Solving the energy equation yields the temperature distribution. Due to the nonlinear nature of the equation, multiple iterations are required. The number of sub-iterations is set within each time step, for example, 5 sub-iterations. The temperature field is updated in each sub-iteration until the convergence condition is met.

[0047] Convergence criteria include velocity residuals and temperature residuals. The velocity residual must be less than 10.-4 Temperature residual requirement is less than 10 -5 The computation is considered convergent when the residuals of all governing equations meet the requirements.

[0048] Outputs three-dimensional field data for multiple future time points. At each time point, complete temperature, velocity, and pressure field data are output. The temperature field data includes the temperature value for each grid node, the velocity field data includes the velocity vector for each grid node, and the pressure field data includes the pressure value for each grid node. These data are stored in chronological order, forming a three-dimensional thermal field evolution sequence.

[0049] The time step setting needs to consider numerical stability. According to the Courant number criterion, the time step should not be too large. A time step of 0.1 seconds is typically set. The prediction time range is determined based on actual needs; for example, predicting the thermal field evolution over the next 10 minutes, outputting data at 600 time points.

[0050] The entire solution process is accelerated using parallel computing techniques. The computational domain is divided into multiple sub-regions, each assigned to a different computational core for processing. Data exchange between sub-regions occurs through a message passing interface. Parallel computing significantly improves computational efficiency, enabling real-time prediction.

[0051] To ensure computational accuracy, mesh independence verification is also required. By comparing the computational results under different mesh sizes, a suitable mesh density is determined. If further mesh refinement has minimal impact on the results, the mesh density is considered sufficient. Simultaneously, time step independence verification is also necessary to ensure that the time step is sufficiently small.

[0052] Intermediate results are saved during the calculation process for later analysis and debugging. If the calculation diverges, it can restart from the last converged time step. The system also has an automatic monitoring mechanism that automatically adjusts the solution parameters or restarts the calculation when an anomaly is detected.

[0053] In another preferred embodiment of the present invention, the specific process for identifying the heat retention region in the heat retention analysis module is as follows: First, temperature field data from multiple consecutive time steps are extracted from the 3D thermal evolution sequence. The number of time steps extracted is determined based on the analysis requirements, typically selecting data from the most recent 30 time steps. The temperature field data for each time step includes the temperature values ​​of all grid nodes, arranged in order of grid number. During data extraction, the continuity of the time series must be ensured. If data is missing from a particular time step, linear interpolation is used to supplement the data for that time step.

[0054] Next, the rate of temperature change over time is calculated for each grid point. For each grid point, the temperature values ​​for 10 consecutive time steps are taken, and a trend line of temperature change over time is fitted using the least squares method. The rate of temperature change is defined as the slope of this trend line, in degrees Celsius per second. For example, if the temperature at a grid point rises from 25 degrees Celsius to 26 degrees Celsius in 10 seconds, its rate of temperature change is 0.1 degrees Celsius per second. After the calculation, a temperature rate of change value is obtained for each grid point.

[0055] Simultaneously, the temperature gradient between each grid point and its adjacent grid points is calculated. The temperature gradient calculation uses the central difference method. For each grid point, its six adjacent grid points are identified, and the temperature differences in the X, Y, and Z directions are calculated separately, then divided by the grid spacing. The temperature gradient components in the three directions are combined to obtain the temperature gradient value of that grid point, in degrees Celsius per meter. For example, if the temperature difference between a grid point and its adjacent grid points in the X direction is 0.5 degrees Celsius, and the grid spacing is 0.2 meters, then the temperature gradient component in the X direction is 2.5 degrees Celsius per meter.

[0056] Set a rate of change threshold and a gradient threshold. The rate of change threshold is typically set to 0.05 degrees Celsius per second, representing regions where temperature changes are relatively slow. The gradient threshold is typically set to 1.0 degrees Celsius per meter, representing regions where temperature distribution is relatively uniform. Compare the temperature change rate of each grid point with the rate of change threshold, and compare the temperature gradient value with the gradient threshold. When the temperature change rate of a grid point is lower than the rate of change threshold and the temperature gradient value is lower than the gradient threshold, that grid point is marked as a candidate point.

[0057] Spatial clustering analysis is performed on the candidate points. A density-based clustering algorithm is used to group spatially adjacent candidate points into the same category. Adjacent grid points are defined as those whose distance in the X, Y, and Z directions does not exceed one grid spacing. During the clustering process, an unvisited candidate point is randomly selected as the starting point, and all density-reachable candidate points are searched to form a cluster. This process is repeated until all candidate points have been visited. Each cluster represents a candidate region.

[0058] Calculate the geometric characteristics of each candidate region. For each candidate region, first calculate its volume, which is equal to the number of grid cells in the region multiplied by the volume of a single grid cell. For example, if a candidate region contains 50 grid cells, and each grid cell has a volume of 0.008 cubic meters, then the region's volume is 0.4 cubic meters. Next, calculate the region's average temperature by taking the arithmetic mean of the temperature values ​​at all grid points in the region. Simultaneously, record the region's maximum temperature value, which is the highest temperature value among all grid points.

[0059] Set volume and temperature thresholds. The volume threshold is typically set to 0.2 cubic meters, representing a region of a certain size. The temperature threshold is typically set to 28 degrees Celsius, representing a region with higher temperatures. Compare the volume of each candidate region with the volume threshold, and compare the average temperature with the temperature threshold. When a candidate region has a volume greater than the volume threshold and an average temperature higher than the temperature threshold, the candidate region is marked as a thermal stagnation region.

[0060] Finally, the spatial extent data of the thermal confinement region is output. For each thermal confinement region, a list of the numbers of all grid cells it contains is recorded. The geometric center coordinates of the region are also calculated, obtained by averaging the coordinates of all grid points. The output data also includes attribute data such as volume, average temperature, and maximum temperature for each thermal confinement region. This data is stored in order of region numbering, forming a complete thermal confinement region analysis report.

[0061] The entire identification process employs an iterative optimization method. After each identification, the result is saved and compared with the results of subsequent time steps. If a region is identified as a heat stagnation area for multiple consecutive time steps, the warning level for that region is increased. Simultaneously, the system dynamically adjusts various threshold parameters based on historical identification results to make the identification results more accurate and reliable.

[0062] In another preferred embodiment of the present invention, the specific process of generating the guide vane angle control parameters and calculating the task migration parameters in the flow control module is as follows: First, the spatial distribution characteristics of the heat retention area are analyzed. The spatial range data output by the heat retention area analysis module is read, and the geometric characteristic parameters of each heat retention area are extracted. Geometric features are characterized by calculating the area's aspect ratio and surface area-to-volume ratio. Spatial location features are recorded as the three-dimensional coordinates of the area's center point. Volume data is directly taken from the area's volume calculation results. Simultaneously, the relative positional relationship between each area and each guide vane is calculated, including distance and azimuth.

[0063] The deflector control mapping table is queried based on the characteristics of the heat retention area. The row index of the mapping table is the discretized value of the center coordinates of the heat retention area, and the column index is the discretized value of the area volume. For example, the machine room space is divided into a 1-meter square cubic grid, and the area volume is discretized at 0.1 cubic meter intervals. During the query, the actual area center coordinates are first mapped to the nearest grid point, and the actual volume is mapped to the nearest volume interval. Then, the corresponding initial angle setting is found in the mapping table. The initial angle setting contains the suggested deflection angles for all relevant deflectors.

[0064] The initial angle setting is corrected by combining real-time airflow velocity data and temperature gradient data. Current airflow velocity field data is obtained from the airflow state acquisition module, and current temperature gradient field data is obtained from the thermal field evolution prediction module. The correction process considers the following factors: appropriately increasing the deflection angle of the guide vane in areas with low airflow velocity, and adjusting the direction of the guide vane in areas with large temperature gradients to enhance local heat transfer. The correction amount is calculated using empirical formulas; for example, for every 0.1 m / s decrease in airflow velocity, the guide vane angle increases by 1 degree. Finally, guide vane angle control parameters containing specific angle values ​​and execution timestamps are generated. Identify the computational tasks running on the computing nodes within the hot stifling region. Based on the spatial extent of the hot stifling region, locate all computing nodes within that region. Obtain a list of tasks running on these nodes using the computational feature acquisition module. Each task record includes information such as task identifier, start time, elapsed running time, and resource utilization. Simultaneously, extract computational feature data for these tasks from the computational feature set, including instruction type distribution, data access patterns, and floating-point operation intensity.

[0065] Task migration priorities are determined based on the floating-point operation intensity in the computational feature data. Tasks are sorted from highest to lowest floating-point operation intensity, with tasks having higher floating-point operation intensity being migrated first. For example, tasks with a floating-point operation intensity greater than 10 are marked as high priority, tasks with a floating-point operation intensity between 5 and 10 are marked as medium priority, and tasks with a floating-point operation intensity less than 5 are marked as low priority. The running time of the tasks is also considered, with tasks with shorter running times being migrated first.

[0066] When selecting target computing nodes, temperature distribution and resource utilization requirements must be met. Current temperature field data is obtained from the thermal field evolution prediction module, and computing nodes within regions with temperatures below 26 degrees Celsius are selected. Node resource utilization data is obtained from the computing feature acquisition module, and nodes with CPU utilization below 60% and memory utilization below 70% are selected. When multiple nodes meet the requirements, the node with lower resource utilization is prioritized.

[0067] Generate computation task migration parameters. These parameters include the task identifier, source node internet address, target node internet address, migration priority, and execution time window. For high-priority tasks, a shorter execution time window is set, requiring migration to complete within 1 minute. For medium- and low-priority tasks, a longer time window is set, allowing migration to complete within 5 minutes.

[0068] The process of constructing the deflector control mapping table is as follows: A parametric mesh was created based on a 3D model of the data center. The cell size of the parametric mesh was consistent with the accuracy of thermal congestion zone identification, both being 0.1 meters. The mesh covered the entire data center space, including all server racks, aisles, and ventilation facilities. Each mesh cell recorded its spatial coordinates and type attributes, including solid region, fluid region, and baffle region.

[0069] Flow and temperature field distributions under different combinations of guide vane angles were obtained through computational fluid dynamics simulation. Simulation settings included: inlet boundary conditions set according to the actual air supply parameters of the air conditioning system, outlet boundary conditions set to a pressure outlet, and heat source distribution set according to the actual power consumption of the server. The guide vane angle ranged from -45 degrees to +45 degrees, with simulations performed at 5-degree intervals. After each angle combination simulation was completed, complete flow and temperature field data were recorded.

[0070] A database was established to correlate the angles of the deflector with the effectiveness of eliminating heat stagnation areas. The database records the changes in volume and temperature of each heat stagnation area, as well as the degree of improvement in airflow organization, for each combination of deflector angles. The elimination effect is quantified by the percentage reduction in heat stagnation area volume and the average temperature decrease.

[0071] A supervised learning method was used to train the angle recommendation model. Training samples were derived from computational fluid dynamics simulations. Input features included the location coordinates, volume, and temperature gradient of the thermal stagnation region. The output label was the optimal combination of guide vane angles. A gradient boosting decision tree algorithm was used for training, with a maximum tree depth of 10 and a learning rate of 0.1. Five-fold cross-validation was employed to evaluate model performance during training.

[0072] Data on the control effect of the deflector was collected during actual operation. After each deflector adjustment, temperature and airflow velocity data were recorded before and after the adjustment. Particular attention was paid to the volume and temperature changes in the heat retention area after adjustment. The deviation between the actual deflector angle and the set value was also recorded.

[0073] An online update mechanism based on reinforcement learning was established. The deflector control was modeled as a Markov decision process, with states including thermal stagnation region characteristics and environmental parameters, actions including deflector angle setting, and a reward function based on thermal stagnation region elimination effectiveness and energy consumption indicators. A deep Q-learning algorithm was used to continuously optimize the angle recommendation strategy; the network structure consisted of three fully connected layers, each with 128 neurons.

[0074] Periodically verify the performance of the updated deflector control mapping table under actual operating conditions. A comprehensive verification should be conducted weekly, selecting typical operating conditions to test the control effect of the new mapping table. Verification metrics include heat retention area elimination rate, energy consumption change rate, and control stability. When the performance metrics are superior to the old mapping table, replace the old mapping table with the new one.

[0075] The entire flow control module operates using a layered decision-making architecture. The bottom layer is responsible for real-time data acquisition and processing, the middle layer executes control algorithms to generate control parameters, and the top layer performs strategy optimization and performance evaluation. Each layer exchanges data and coordinates with the others through clearly defined interfaces to ensure the stability and reliability of the control system.

[0076] In another preferred embodiment of the present invention, the specific process of transmitting instructions to the execution device in the instruction execution module is as follows: First, the guide vane angle control parameters are processed. The angle values ​​are converted into a stepper motor drive protocol format, which contains a 32-bit data frame. 16 bits represent the target angle, 8 bits represent the rotation speed, and the remaining 8 bits contain angle calibration and error compensation instructions. The angle calibration instruction corrects the motor's zero-position deviation, and the error compensation instruction eliminates transmission backlash errors. A cyclic redundancy check (CRC) code is added to the protocol data frame; the check code is 16 bits long.

[0077] Simultaneously, computational task migration parameters are processed. These parameters are encapsulated into distributed system communication protocol data packets. The packet header includes a version number, timestamp, and sequence number. The protocol body contains task state saving instructions and task state restoration instructions. The task state saving instructions guide the source node to save the task's runtime environment, including register states, memory page tables, and file descriptors. The task state restoration instructions guide the target node to rebuild the task's runtime environment. An MD5 checksum is appended to the end of the data packet.

[0078] The flow deflector control signals are transmitted via real-time industrial Ethernet. The network employs time-division multiple access (TDMA) to assign the highest transmission priority to control signals. Each network frame contains a 64-byte payload, with a transmission period of 1 millisecond. The network switch is configured with a strict priority queue to ensure that control signals are forwarded preferentially. The transmission path uses twisted-pair cabling, and the interface conforms to the IEEE 802.3 standard.

[0079] Task migration commands are transmitted via the data center network. The network architecture uses a leaf-spine topology and is configured with equal-cost multipath routing. Each command is transmitted simultaneously via two independent physical paths, and the receiving end processes them on a first-come, first-served basis. The network links use the 25 Gigabit Ethernet standard, with a transmission latency of less than 100 microseconds. Switches exchange routing information via Border Gateway Protocol (BGP).

[0080] Establish a monitoring loop for command execution status. An absolute encoder is installed at the guide plate end to provide real-time feedback on the motor rotation angle; the feedback data is transmitted back to the control terminal via an I / O link. A task migration monitor is deployed on the computing node to report the task status saving and recovery progress in real time. The monitoring data is sampled at 100 Hz and transmitted through an independent monitoring network.

[0081] A retransmission mechanism is designed to handle transmission anomalies. The flow guide control command sets a maximum of 3 retransmissions, while the task migration command sets a maximum of 5 retransmissions. Retransmission is triggered when a transmission timeout or verification failure occurs. The retransmission interval uses an exponential backoff algorithm, with an initial interval of 10 milliseconds and a maximum interval of 100 milliseconds. The retransmission counter is reset to zero after each successful transmission.

[0082] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A control system for a high-efficiency energy-saving data center computer room, characterized in that, include: The airflow state acquisition module is used to collect airflow trajectory data and equipment surface temperature distribution data in the computer room space, forming an airflow thermal field state set that includes velocity field and temperature field; The computational feature acquisition module is used to collect computational instruction type data, data access mode data, and floating-point operation intensity data during server operation, forming a set of computational features that reflect the characteristics of computational load. The thermal field evolution prediction module is used to align airflow trajectory data with equipment surface temperature distribution data in time and space, establish an unsteady airflow-thermal field coupling equation set, and convert floating-point operation intensity data into heat source terms to input the equation set. Through iterative solution, a three-dimensional thermal field evolution sequence for future time periods is obtained. The thermal retention analysis module is used to analyze the temperature gradient distribution in the three-dimensional thermal field evolution sequence, identify areas where the temperature gradient is continuously below a set threshold, and mark them as thermal retention areas. The flow control module is used to generate flow deflector angle control parameters and calculation task migration parameters based on the spatial distribution characteristics of the heat retention area. The instruction execution module is used to convert the guide vane angle control parameters into guide vane drive signals, convert the calculation task migration parameters into task reassignment instructions, and transmit them to the execution device.

2. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, The specific process of collecting airflow trajectory data and equipment surface temperature distribution data in the airflow state acquisition module is as follows: Multiple measurement points are set up in the computer room, the coordinate position of each measurement point is recorded, the air flow velocity and temperature value of each measurement point are collected, the air flow velocity is converted into velocity vector data, and the temperature value is mapped to the coordinate position of the corresponding measurement point. The velocity vector data and temperature values ​​are organized in chronological order. Data with inconsistent timestamps are time-aligned. Missing data is interpolated using values ​​from adjacent time points. The processed velocity vector data and temperature values ​​are stored as a three-dimensional array containing time, space, and data dimensions.

3. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, The specific process of acquiring data from the server during runtime in the computational feature acquisition module is as follows: The number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions are counted to form computation instruction type data; cache access frequency, memory read / write frequency, and storage input / output operation frequency are recorded to form data access pattern data. The ratio of the number of floating-point operations to the number of data accesses per unit time is used to form floating-point operation intensity data. The collected data is classified according to the processor core, and a collection time stamp is attached to each data item. A hierarchical data table is used to store the set of calculated features. The first layer of the hierarchical data table identifies the processor number, and the second layer of the hierarchical data table stores the time series of various feature data.

4. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, The specific process of obtaining the three-dimensional thermal field evolution sequence in the thermal field evolution prediction module is as follows: Airflow trajectory data is resampled to a non-uniform computing grid, wherein the grid spacing of the non-uniform computing grid is less than a set threshold in densely populated areas of equipment and greater than a set threshold in sparsely populated areas of equipment. Apply heat flux boundary conditions based on the measured temperature distribution at the boundary of the computational domain; An unsteady coupled equation system incorporating buoyancy and turbulence effects is constructed. A dynamic heat source term derived from floating-point arithmetic intensity data is introduced into the unsteady coupled equation system. A step-by-step solution strategy is adopted to first solve the flow field equations, and then input the flow field solution results into the temperature field equations. Multiple sub-iteration calculations are performed within each time step to output three-dimensional temperature field data, three-dimensional velocity field data, and three-dimensional pressure field data at multiple future time points.

5. The control system for a high-efficiency energy-saving data center room according to claim 4, characterized in that, The specific process of deriving the dynamic heat source term is as follows: Establish the correspondence between computing instruction type and heat generation per unit time, determine the heat-to-work conversion coefficient of floating-point arithmetic unit based on processor architecture parameters, collect processor operating frequency and voltage data, and calculate instantaneous heat generation power in combination with the heat-to-work conversion coefficient; Based on the spatial coordinates of the processor, the instantaneous heat generation power is distributed to the corresponding grid cells. The thermal resistance parameters of the heat sink and the thermal conductivity of the interface material are read, and the distribution characteristics of the heat source in the grid cells are calculated to form heat source distribution data containing spatial location attributes and time variation characteristics. The heat source distribution data is used as a dynamic heat source term input into the unsteady coupled equation system.

6. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, The specific process for identifying the heat retention region in the heat retention analysis module is as follows: Temperature field data for multiple consecutive time steps are extracted from the three-dimensional thermal field evolution sequence. The temperature change rate of each grid point in the time dimension is calculated, and the temperature gradient value between each grid point and its neighboring grid points is calculated. Grid points with a temperature change rate lower than the change rate threshold and a temperature gradient value lower than the gradient threshold are marked as candidate points. The candidate points are spatially clustered to form continuous regions. The volume, average temperature and maximum temperature of each continuous region are calculated. Continuous regions with a volume greater than the volume threshold and an average temperature higher than the temperature threshold are marked as thermal stagnation regions. The spatial range data of the thermal stagnation regions are output.

7. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, In the flow control module, the specific process of generating the flow deflector angle control parameters and calculating the task migration parameters is as follows: The spatial distribution characteristics of the heat retention area are analyzed, including the region's geometry, spatial location, and volume data. Based on the characteristics of the heat retention area, the deflector control mapping table is queried to obtain the initial angle setting. The initial angle setting is then corrected by combining real-time airflow velocity data and temperature gradient data to generate the deflector angle control parameters. Identify the computing tasks running on computing nodes within the thermal stagnation region and obtain the computing characteristic data of the tasks; The task migration priority is determined based on the floating-point operation intensity in the computational feature data. Target computing nodes with satisfactory temperature distribution and resource utilization below the threshold are selected, and computational task migration parameters containing task identifier, source node address, and target node address are generated.

8. The control system for a high-efficiency energy-saving data center room according to claim 7, characterized in that, The process of constructing the deflector control mapping table is as follows: A parametric mesh is established based on the 3D model of the computer room. The mesh cell size is consistent with the identification accuracy of the heat retention area. The flow field distribution and temperature field distribution under different combinations of guide vane angles are obtained through computational fluid dynamics simulation. A database of the correspondence between guide vane angle and heat retention area elimination effect is established. A supervised learning method is used to train the angle recommendation model. The training features include the location, volume, and temperature gradient of the thermal stagnation region. During actual operation, the control effect data of the deflector is collected, including temperature changes and airflow organization changes before and after angle setting. An online update mechanism based on reinforcement learning is established. The angle recommendation strategy is optimized according to the control effect data. The performance of the updated deflector control mapping table under actual operating conditions is periodically verified.

9. The control system for a high-efficiency energy-saving data center room according to claim 1, characterized in that, In the instruction execution module, the specific process of transmitting instructions to the execution device is as follows: The guide vane angle parameters are encoded into a stepper motor drive protocol that includes angle calibration instructions and error compensation instructions; the computation task migration parameters are encapsulated into a distributed system communication protocol that includes task state saving instructions and task state recovery instructions. By employing a priority scheduling mechanism to transmit the flow guide plate control signal in real-time via industrial Ethernet, and by using a data center network with redundant links to transmit the task migration instructions, a monitoring loop for the instruction execution status that collects flow guide plate position feedback and task migration progress in real time is established. A retransmission mechanism is designed to set an upper limit for the number of retransmissions based on the instruction type.

Images