A computing power center energy consumption monitoring and intelligent energy-saving scheduling method

Abstract

The present application belongs to the technical field of industrial monitoring and dispatching, and particularly relates to a computing power center energy consumption monitoring and intelligent energy-saving dispatching method, which comprises the following steps: acquiring and preprocessing IT power consumption, cooling equipment electric power, temperature, flow and task data of each cooling partition, determining a thermal influence neighborhood and a periodic variation, online identifying a thermal influence transmission coefficient and a heat conduction delay period, and generating a spatial thermal influence transmission coefficient matrix and a delay matrix; according to a communication topology, a cooling equipment state and an available heat removal capacity, candidate cooling partitions are screened, optimal deployment positions are determined in combination with expected cooling energy consumption of tasks, and cooling equipment electric power target values of each cooling partition are generated; according to cooling equipment electric power change prediction error, parameters are updated, and target values are adjusted according to an air inlet temperature deviation, so that closed-loop correction of task deployment and partition cooling control is realized, and unnecessary cooling energy consumption is reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial monitoring and scheduling technology. More specifically, this invention relates to a method for energy consumption monitoring and intelligent energy-saving scheduling in computing centers. Background Technology

[0002] In the field of energy consumption monitoring and intelligent energy-saving scheduling in computing centers, existing systems typically rely on temperature sensor networks and cooling system controllers. Intake and return air temperature measurement points are set up in server racks or cooling areas to collect IT power consumption, ambient temperature, and cooling equipment operating data. Cooling control generally employs a uniform static temperature threshold or predicts future energy consumption based on historical average loads, uniformly increasing the operating power of cooling equipment when temperatures exceed limits or predicted load increases.

[0003] This type of method has simple control logic, low implementation cost, and is easy to integrate into data center management systems. Under conditions of small data center size, stable task operation, uniform airflow organization, and independent thermal environment of each rack, rack temperature mainly varies with the IT power consumption of the area. Using a uniform threshold or historical load for control can suppress local overheating. Some methods can also increase cooling capacity in advance based on task resource request information.

[0004] However, in large-scale heterogeneous GPU clusters, cross-location thermal effects exist between different locations due to heat diffusion and airflow. Changes in IT power consumption at any location not only alter the cooling requirements of that location but may also delay the impact on the power consumption of cooling equipment in other locations. Existing methods typically treat each location as an independent cooling unit, primarily comparing the resource reserves or cooling capacity of candidate locations during task deployment, which fails to reflect the impact of task deployment on the overall data center cooling load. With changes in task start / stop, airflow organization, and equipment status, deployment judgments and cooling control based on fixed thresholds or historical loads are prone to deviating from the current operating state, causing the task deployment location and corresponding cooling control results to fail to match the cross-location thermal effect relationship and increasing unnecessary cooling energy consumption. Summary of the Invention

[0005] To address the technical problem of existing technologies' inability to coordinate task deployment and zoned cooling control based on cross-location thermal impact relationships, this invention provides a method for monitoring and intelligently scheduling energy consumption in computing centers, comprising:

[0006] Obtain operational status data for each cooling zone within the computing center, as well as the expected maximum IT power consumption and estimated runtime of tasks to be deployed;

[0007] Based on the operating status data of each cooling zone, the heat influence transfer relationship and heat conduction delay relationship between each cooling zone are identified, and the spatial heat influence transfer coefficient matrix and delay matrix are constructed.

[0008] Based on the spatial heat effect transfer coefficient matrix, delay matrix, and communication topology requirements of the task to be deployed, a set of candidate cooling partitions that meet the deployment conditions is selected, and the delay weight of each candidate cooling partition to other cooling partitions is determined.

[0009] Combining latency weights, spatial thermal impact transfer coefficient matrix, and expected maximum IT operating power consumption, the expected cooling energy consumption of the entire data center is calculated when the task to be deployed is deployed to each candidate cooling partition. The candidate cooling partition with the minimum expected cooling energy consumption of the task is determined as the optimal deployment location.

[0010] Based on the expected increase in cooling energy consumption corresponding to the optimal deployment location, update the target power value of the cooling equipment in each cooling zone to schedule the tasks to be deployed and control the cooling equipment.

[0011] Preferably, acquiring the operating status data of each cooling zone within the computing center includes:

[0012] The system collects real-time power consumption, temperature, cooling medium flow rate, and cooling equipment power of each cooling zone; it performs periodic segmentation, timestamp alignment, outlier removal, and missing value completion on the collected raw data to generate a standard time series with valid data labels.

[0013] Preferably, identifying the heat transfer relationship and heat conduction delay relationship between cooling zones based on the operating status data of each cooling zone includes:

[0014] Based on standard time series, the power consumption change and cooling equipment power change of each cooling zone are obtained; the thermal influence neighborhood of each cooling zone is determined according to the cold aisle layout, hot aisle layout and airflow organization direction of the computer room; using a recursive least squares algorithm with forgetting factor and regularization term, the thermal influence transfer coefficient and heat conduction delay period of each source cooling zone to the affected cooling zone are identified according to the power consumption change and cooling equipment power change of each cooling zone in the thermal influence neighborhood.

[0015] Preferably, the power consumption change of each cooling zone within the heat-affected zone satisfies the relationship between the power consumption change of the cooling equipment and the following formula:

[0016] ;

[0017] In the formula, Indicates period Internal cooling partition The change in electrical power of the cooling equipment; Cooling partition For cooling zones The heat conduction delay period; Cooling partition For cooling zones The heat transfer coefficient; Indicates the affected cooling zone number; Indicates the cooling zone number of the affected source; Indicates the current cycle number; Indicates the heat-affected neighborhood; Indicates period Internal cooling partition The change in IT power consumption.

[0018] Preferably, the construction of the space thermal effect transmission coefficient matrix and the delay matrix includes:

[0019] Arrange the heat transfer coefficients in the order that rows represent affected cooling zones and columns represent affected cooling zones to generate a spatial heat transfer coefficient matrix; arrange the heat conduction delay periods in the order that rows represent affected cooling zones and columns represent affected cooling zones to generate a delay matrix.

[0020] Preferably, the step of selecting a set of candidate cooling partitions that meet the deployment conditions based on the spatial thermal effect transfer coefficient matrix, the delay matrix, and the communication topology requirements of the task to be deployed, and determining the delay weight of each candidate cooling partition relative to other cooling partitions, includes:

[0021] Cooling zones that simultaneously meet the network communication hop limit, normal operating status of cooling equipment, and sufficient heat removal capability are added to the candidate cooling zone set. For each candidate cooling zone in the candidate cooling zone set, the delay weight of the candidate cooling zone to other cooling zones is calculated based on the heat conduction delay period and its maximum value in the delay matrix.

[0022] Preferably, the delay weight satisfies the expression:

[0023] ;

[0024] In the formula, Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat conduction delay period; This represents the maximum effective heat conduction delay period in the delay matrix; Indicates the affected cooling zone number; Indicates the candidate cooling zone number.

[0025] Preferably, the optimal deployment location satisfies the expression:

[0026] ;

[0027] In the formula, Indicates tasks to be deployed The optimal deployment location; Indicates tasks to be deployed The set of candidate cooling partitions; Indicates tasks to be deployed The estimated runtime; Indicates tasks to be deployed The expected maximum IT operating power consumption; Indicates the total number of cooling zones; Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat transfer coefficient; This indicates a candidate cooling zone that minimizes the calculation result within the parentheses. To deploy tasks Deploy to candidate cooling partitions The expected cooling energy consumption during the mission.

[0028] Preferably, updating the target power value of the cooling equipment in each cooling zone based on the expected cooling energy consumption increment corresponding to the optimal deployment location includes:

[0029] For each cooling zone, the expected maximum IT operating power consumption is multiplied by the delay weight and heat transfer coefficient of the optimal deployment location for that cooling zone to obtain the initial cooling equipment power increment;

[0030] The current baseline value of the cooling equipment's electrical power is added to the initial increment of the cooling equipment's electrical power to obtain the target value of the cooling equipment's electrical power for the corresponding cooling zone.

[0031] Preferably, after scheduling the tasks to be deployed and controlling the cooling equipment, the method further includes:

[0032] The actual power variation of the cooling equipment in each cooling zone is collected, and the prediction error of the power variation is calculated. Based on the prediction error of the power variation, the parameters in the spatial heat influence transfer coefficient matrix and the delay matrix are updated online. The inlet air temperature deviation of each cooling zone is obtained, and the target power of the cooling equipment in each cooling zone is dynamically adjusted based on the inlet air temperature deviation.

[0033] The beneficial effects of this invention are as follows: This invention performs periodic segmentation, time alignment, anomaly handling, missing data completion, and validity assessment on computing center operation data and task data, ensuring that subsequent calculations use data with consistent time and reliable status; this invention determines the thermal impact neighborhood and continuously updates the thermal impact transfer coefficient and heat conduction delay period based on valid operation data, allowing the thermal impact relationship between different cooling zones to adjust with changes in task load, airflow organization, and cooling equipment operating status; this invention combines communication topology, cooling equipment status, and available heat removal capabilities to screen candidate cooling zones, avoiding the deployment of tasks to locations with unsatisfactory communication conditions or insufficient cooling capacity; this invention compares task deployment... The invention addresses the expected cooling energy consumption of the entire data center when deployed to different candidate cooling zones, ensuring that the optimal deployment location balances task operation requirements with the cooling burden of each cooling zone. It adds the increased cooling power of the cooling equipment caused by the task to the current cooling power of each cooling zone, forming a target value for the cooling equipment power that can be executed by the corresponding cooling system controller. The invention updates the heat transfer relationship based on the predicted error of cooling equipment power changes and adjusts the target value of cooling equipment power based on the inlet air temperature deviation. Simultaneously, it eliminates interference from temperature control actions on parameter updates, ensuring that task deployment results and zone cooling control continuously correspond to the current operating state of the computing center, reducing unnecessary cooling energy consumption. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a computing center energy consumption monitoring and intelligent energy-saving scheduling method according to the present invention. Detailed Implementation

[0035] 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, not all, of the embodiments of the present invention. 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.

[0036] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] This invention discloses a method for monitoring energy consumption and intelligent energy-saving scheduling in computing centers, referring to... Figure 1 This includes steps S1-S7:

[0038] S1: Obtain the computing center's operational data and task data, and preprocess the operational data to obtain a standard time series sequence.

[0039] It should be noted that in this embodiment, the computing center is divided into multiple cooling zones. Each cooling zone has an independent IT power consumption statistical boundary and a cooling equipment power measurement boundary, and includes at least one server rack. When a cooling device serves multiple server racks simultaneously, all server racks served by that cooling device are assigned to the same cooling zone. When a cooling zone includes multiple cooling devices, the sum of the power of the multiple cooling devices is taken as the cooling equipment power of that cooling zone.

[0040] Real-time IT power consumption of each cooling zone is acquired via a smart PDU. Inlet and outlet air temperatures of each cooling zone are acquired via a temperature sensor array. Cooling medium flow rates of the corresponding cooling equipment in each cooling zone are acquired via flow meters. The operating status and power consumption of the cooling equipment are acquired via the cooling system controller interface. The difference between the inlet and outlet air temperatures is used to verify whether temperature data changes correspond to current IT power consumption changes. Cooling medium flow rates are used to verify whether the cooling equipment is executing cooling control commands. When a cooling control command changes, if the cooling medium flow rate does not change for two consecutive uniform cycles, the corresponding cooling equipment is marked as abnormal; otherwise, the corresponding cooling equipment is marked as having a normal cooling medium flow rate verification result. In other embodiments, the number of consecutive cycles can be set according to the control command transmission time and the cooling equipment response time.

[0041] Simultaneously, obtain the queued task list, GPU request quantity, estimated runtime, and inter-node communication topology requirements from the Slurm scheduler interface or Kubernetes scheduler interface and the task submission script. Then, retrieve the tasks to be deployed. The expected maximum IT operating power consumption is denoted as When the task submission information includes the maximum IT operating power consumption, this maximum IT operating power consumption is used as the declared power consumption value. The number of GPUs requested is multiplied by the thermal design power (TDP) of the corresponding GPU model, and the reserved maximum power consumption of the CPU, memory, and other components in the task-occupied node is added to obtain the estimated hardware power consumption value. When both the declared power consumption value and the estimated hardware power consumption value are available, the larger of the two is used as the expected maximum IT operating power consumption, denoted as [missing value]. When only one power consumption value can be obtained, the obtainable power consumption value is recorded as the expected maximum IT operating power consumption. Tasks to be deployed The estimated runtime is converted to hours and recorded as Used for subsequent calculations of tasks to be deployed. Cooling energy consumption during the estimated runtime; determine tasks to be deployed based on inter-node communication topology requirements. Communication hop count requirements.

[0042] All real-time collected physical quantities are segmented according to a preset uniform period. Within each uniform period, the periodic average value of each physical quantity is calculated to form an equally spaced original time sequence. In this embodiment, the uniform period is set to 30 seconds. In other embodiments, the uniform period can be set to 10 to 60 seconds, depending on the sensor sampling rate, the rate of change of task power consumption, and the response time of the cooling device.

[0043] Using a uniform period as the time grid step size, asynchronous data generated by different sensors and system interfaces are aligned to the same time axis. For time-series signals with inconsistent sampling rates, linear interpolation or nearest neighbor alignment methods are used to generate data on the corresponding time grid.

[0044] Erroneous readings are identified based on the reasonable range of values ​​for each physical quantity and the historical statistical characteristics of each measuring point. Examples include negative power readings and readings where the temperature change exceeds a preset temperature gradient within a single uniform period. These erroneous readings are then removed from the original time series. In this embodiment, the preset temperature gradient is 2 degrees Celsius per uniform period. In other embodiments, the preset temperature gradient can be set based on the accuracy of the temperature sensor, the normal temperature rise rate of the server, and the response time of the cooling equipment.

[0045] For data missing data lasting no more than three uniform periods, linear interpolation is performed using the valid values ​​before and after the missing location, or the data change trend of the most recent five valid periods is used for filling. For data missing data lasting more than three uniform periods, the corresponding interval is marked as an unreliable interval. In other embodiments, the short-term missing data limit and the filling window length can be set according to the duration of communication packet loss, sensor sampling rate, and the allowable number of consecutive missing samples. Here, three uniform periods is the short-term missing data limit set in this embodiment, and five valid periods is the filling window length set in this embodiment. In other embodiments, implementers can set the short-term missing data limit based on the duration of communication packet loss, sensor sampling rate, and the allowable number of consecutive missing samples, and set the filling window length based on the rate of change of the physical quantity and the amount of valid data.

[0046] Data validity markers are set for each cooling zone and each uniform cycle. Parameter updates are only performed using the corresponding data when the IT power consumption, cooling equipment power consumption, and data within the corresponding latency period are all valid, and the cooling equipment is not in a fault or maintenance state. Power is uniformly converted to kilowatts, temperature to degrees Celsius, time to seconds, estimated task runtime to hours, and flow rate to the standard flow rate unit used by the cooling system, thereby forming standard time-series data.

[0047] S2: Determine the heat impact neighborhood of each cooling zone based on the airflow path in the computer room, and calculate the changes in IT power consumption and the changes in the electrical power of cooling equipment.

[0048] It should be noted that the power consumption change of any cooling zone will only change the cooling demand of other cooling zones if the airflow path can reach them. In order to prevent the combination of zones without actual airflow transmission paths from participating in the parameter calculation, this invention determines the thermal influence neighborhood of each cooling zone.

[0049] Specifically, based on the cold aisle layout, hot aisle layout, rack row arrangement, and airflow direction of the computer room, each affected cooling zone is... Set the heat-affected neighborhood The heat-affected neighborhood Including cooling zones itself, and cooling zones Adjacent cooling zones in the same column, and cooling zones Cooling zones arranged face-to-face and cooling zones located on the same hot aisle return air path. (And cooling zones) Long-distance cooling zones with no airflow path between them do not belong to the thermally affected zone. .

[0050] Cooling partitions IT power consumption changes affect cooling zones The magnitude of the impact of changes in the electrical power of the cooling equipment is denoted as the heat transfer coefficient. Cooling partition The IT power consumption changed until that change began to affect the cooling zone. The uniform number of cycles required for cooling is denoted as the heat conduction delay period. For those not belonging to the heat-affected zone The source of influence is the cooling zone Cooling partition For cooling zones Heat transfer coefficient Set to 0, and do not perform recursive updates on this heat effect transfer coefficient; for those belonging to the heat effect neighborhood Cooling partition Retain the corresponding heat transfer coefficient and heat conduction delay period Among the parameters to be identified, the heat conduction delay period of the cooling zone itself is... Set to 0.

[0051] Get cooling zone In the cycle Average power of internal cooling equipment and cooling zones In the cycle Average IT power consumption within The difference between the average power consumption of the cooling equipment in the current cycle and the average power consumption of the cooling equipment in the previous cycle is taken as the change in power consumption of the cooling equipment. The difference between the average IT power consumption in the current cycle and the average IT power consumption in the previous cycle is taken as the change in IT power consumption. The calculation relationship is as follows:

[0052] ;

[0053] ;

[0054] In the formula, Indicates period Internal cooling partition The change in electrical power of the cooling equipment; Indicates period Internal cooling partition The average power of the cooling equipment; Indicates period Internal cooling partition The average power of the cooling equipment; Indicates period Internal cooling partition The change in IT power consumption; Indicates period Internal cooling partition Average IT power consumption; Indicates period Internal cooling partition Average IT power consumption; Indicates the affected cooling zone number; Indicates the cooling zone number of the affected source; Indicates the current cycle number.

[0055] To prevent sampling fluctuations during periods of stable power consumption from continuously altering the thermal impact transfer coefficient, a power consumption change dead zone threshold is set for each affected cooling zone. :

[0056] ;

[0057] In the formula, Indicates cooling zones The power consumption change dead zone threshold; Indicates cooling zones The standard deviation of IT power consumption variation during periods of stable power consumption, specifically within 20 consecutive uniform cycles of no task startup, no task shutdown, and stable power consumption in the cooling zone. The standard deviation of IT power consumption variation, in other embodiments, the length of the power consumption stability period can be set according to the task start / stop frequency and the basic power consumption fluctuation period.

[0058] Furthermore, when heat affects the surrounding area Skip the cooling zone if the absolute value of the IT power consumption change in all cooling zones does not exceed their respective power consumption change dead zone thresholds. The corresponding parameters are updated in this cycle; when the absolute value of the IT power consumption change of at least one cooling zone exceeds the corresponding power consumption change dead zone threshold, subsequent recursive identification is performed.

[0059] S3: Based on the changes in IT power consumption and cooling equipment power consumption after the delay, identify the heat transfer coefficient and heat conduction delay period online.

[0060] Cooling partition When the power consumption of IT equipment changes, the generated heat reaches the cooling zone through thermal diffusion and airflow. Cooling partition The corresponding cooling equipment then changes its electrical power, therefore, the cooling zones... The change in electrical power of the cooling equipment can be determined by the thermal influence neighborhood. The delay IT power consumption change of each cooling zone within the system is represented by the corresponding heat transfer coefficient, and the cooling zones are constructed accordingly. The formula for the change in electrical power of the cooling equipment is as follows:

[0061] ;

[0062] In the formula, Indicates period Internal cooling partition The change in electrical power of the cooling equipment; Cooling partition For cooling zones The heat conduction delay period indicates the cooling zone. The power consumption changes began to affect the cooling partitions. The uniform number of cycles required for cooling; Cooling partition For cooling zones The heat transfer coefficient represents the cooling zone. For every 1kW increase in IT power consumption, after After a unified cycle, the cooling zones are... Increased power consumption of cooling equipment times; Indicates the affected cooling zone number; Indicates the cooling zone number of the affected source; Indicates the current cycle number; Indicates the heat-affected neighborhood; Indicates period Internal cooling partition The change in IT power consumption.

[0063] Before starting the recursive identification, the thermal influence transfer coefficients of each to be identified are... Initialized to 0, using data from 60 consecutive valid cycles, the cooling zone is... IT power consumption change sequence and cooling partition Cross-correlation analysis was performed on the power variation sequence of the cooling equipment, and the delay period with the largest positive correlation value was selected as the delay period within a uniform delay range of 0 to 10 periods. If multiple values ​​have the same maximum value, the smaller delay period is used. Specifically, 60 consecutive valid periods constitute the cross-correlation data length set in this embodiment, and 10 uniform periods constitute the maximum delay period set in this embodiment. In other embodiments, implementers can set the cross-correlation data length and maximum delay period based on the airflow path length, air supply speed, and cooling equipment response time.

[0064] For each affected cooling zone The heat will affect the surrounding area. The heat transfer coefficients of each influencing source cooling zone in the previous cycle and the heat transfer coefficients of each influencing source cooling zone in the cycle Substituting the changes in IT power consumption within the system into the aforementioned formula for the changes in electrical power of the cooling equipment, we obtain the cooling zone... In the cycle Predicted value of power variation of internal cooling equipment .

[0065] Subtracting the predicted value of the power change of the cooling equipment from the measured value of the power change of the cooling equipment yields the prediction error of the power change of the cooling equipment:

[0066] ;

[0067] In the formula, Indicates period Internal cooling partition The error in predicting changes in the electrical power of cooling equipment; Indicates period Internal cooling partition The measured value of the change in electrical power of the cooling equipment; Indicates cooling zones Predicted value of power variation of cooling equipment.

[0068] For each affected cooling zone According to the heat-affected zone Cooling zones for each influencing source In a fixed order, the heat transfer coefficients obtained in the previous cycle are... Arranged sequentially as a parameter column vector Cooling zones for each influencing source will be configured in the corresponding cycle. The changes in IT power consumption within the input are arranged in the same order as an input column vector. Parameter column vector The heat transfer coefficients and input column vectors in the data The changes in power consumption of each IT component are corresponding one-to-one with the cooling zone number of the influencing source.

[0069] When heat affects the surrounding area The absolute value of the IT power consumption change in at least one cooling zone exceeds the corresponding power consumption change dead zone threshold, and the cooling zone When all valid data labels for the cooling zones involved in the calculation are valid, a recursive least squares method with a forgetting factor and L2 regularization is used, based on the input column vector. Changes in the electrical power of cooling equipment Update the parameter column vector with the recognition results from the previous cycle. The forgetting factor is used to reduce the influence of earlier period data on the current identification result, and the L2 regularization term is used to limit the abnormal increase of the thermal effect transfer coefficient when the input data is correlated. In this embodiment, the forgetting factor is used... Setting the L2 regularization coefficient to 0.99 will adjust the L2 regularization factor. Setting it to 0.11 will change the parameter column vector. The vector is set to 0. When a heat transfer coefficient less than 0 appears after the update, the heat transfer coefficient is set to 0. In other embodiments, the implementer can set the forgetting factor according to the rate of change of operating conditions, and set the L2 regularization coefficient according to the correlation of input data and the degree of parameter fluctuation.

[0070] For the heat conduction delay period During updates, the prediction mean square error (MSE) for the most recent 10 effective periods is calculated using the current delay period, the previous delay period, and the next delay period, respectively, within a range of 0 to 10 uniform periods. The heat conduction delay period is corrected by one uniform period only if the MSE corresponding to an adjacent delay period is reduced by at least 5% compared to the MSE corresponding to the current delay period; otherwise, it remains unchanged. Here, the most recent 10 effective periods are the error comparison period number set in this embodiment, and 5% is the error reduction percentage set in this embodiment. In other embodiments, the implementer can set the error comparison period number and error reduction percentage based on the cooling equipment response fluctuations and the stability of the delay period.

[0071] When heat affects the surrounding area The absolute value of the IT power consumption change in all cooling zones did not exceed their respective power consumption change dead zone thresholds, or the cooling zones If there are valid data zones marked as invalid in the cooling zones involved in the calculation, the parameter column vector and heat conduction delay period are not updated. The data of each unified period is only used for one recursive update, and the role of historical period data is retained by the internal recursive state of the recursive least squares method, so there is no need to reread all historical period data in subsequent periods.

[0072] S4: Generate the spatial heat transfer coefficient matrix and delay matrix according to the fixed order of the affected cooling zone and the source cooling zone.

[0073] It should be noted that step S3 yields the heat influence transfer coefficient and heat conduction delay period corresponding to each cooling zone combination. In order to use the parameters of all cooling zones in the entire computer room simultaneously in the task deployment calculation, this invention generates two matrices in a fixed order of the source cooling zone and the affected cooling zone.

[0074] Specifically, the heat transfer coefficients are arranged in the order of rows representing affected cooling zones and columns representing affected cooling zones to generate a spatial heat transfer coefficient matrix. The heat conduction delay periods are arranged in the order of rows representing the affected cooling zones and columns representing the influencing cooling zones to generate a delay matrix. :

[0075] ;

[0076] ;

[0077] In the formula, Represents the space thermal effect transfer coefficient matrix; Represents the delay matrix; Indicates the total number of cooling zones; This indicates the affected cooling zone number corresponding to the matrix row; This indicates the sequence number of the cooling zone corresponding to the influence source in the matrix column. For combinations of cooling zones that do not belong to the heat influence neighborhood, the corresponding heat influence transfer coefficient is set to 0, and the corresponding heat conduction delay period is not included in the subsequent maximum delay period calculation.

[0078] In a specific embodiment, for a computing center with three zones arranged in a straight line, the total number of cooling zones satisfies Space heat transfer coefficient matrix Represented as:

[0079] ;

[0080] In the formula, the first row , and These represent the heat transfer coefficients of the first, second, and third cooling zones to the first cooling zone, respectively; the second row... , and These represent the heat transfer coefficients of the first, second, and third cooling zones to the second cooling zone, respectively; the third row... , and These represent the heat transfer coefficients from the first cooling zone, the second cooling zone, and the third cooling zone to the third cooling zone, respectively.

[0081] Delay matrix Represented as:

[0082] ;

[0083] In the formula, the first row , and These represent the heat conduction delay periods of the first, second, and third cooling zones to the first cooling zone, respectively; the second row... , and These represent the heat conduction delay periods of the first, second, and third cooling zones to the second cooling zone, respectively; the third row... , and These represent the heat conduction delay periods of the first, second, and third cooling zones to the third cooling zone, respectively.

[0084] S5: Determine candidate cooling zones based on communication topology, cooling equipment status, and available heat removal capabilities, and calculate the delay weight corresponding to each candidate cooling zone.

[0085] It should be noted that the space heat transfer coefficient matrix and delay matrix This can represent the impact of increased task power consumption on the overall cooling equipment power in the data center, but not all cooling zones are suitable for deploying the task to be deployed. Candidate cooling partitions also need to meet communication topology conditions, cooling equipment status conditions, and heat removal capability conditions. Therefore, this invention first filters candidate cooling partitions based on communication topology conditions, cooling equipment status conditions, and heat removal capability conditions, and then uses the delay matrix... Set delay weights for each candidate cooling zone.

[0086] Specifically, establish tasks to be deployed. Candidate cooling partition set A cooling partition is added to the candidate cooling partition set only if it meets all of the following conditions. :

[0087] First, the cooling zone is related to the task to be deployed. The maximum number of communication hops between assigned nodes in the data center network does not exceed a preset communication hop limit. In this embodiment, the preset communication hop limit is set to two hops. In other embodiments, the preset communication hop limit can be set according to task communication latency requirements, network bandwidth, and data center network topology.

[0088] Second, the cooling equipment corresponding to the cooling zone is in normal operation, not in a faulty or maintenance state, and the cooling medium flow rate verification result is normal.

[0089] Third, the existing IT power consumption and tasks to be deployed in this cooling zone. Expected maximum IT operating power consumption The sum of these values ​​does not exceed the available heat removal capacity of the cooling zone after deducting a safety margin from the rated heat removal capacity, which is obtained from the rated cooling capacity on the nameplate of the cooling equipment. In this embodiment, the safety margin is set to 10% of the rated heat removal capacity. In other embodiments, the safety margin can be set according to fluctuations in task power consumption, changes in cooling equipment performance, sensor measurement errors, and equipment reliability requirements.

[0090] If the candidate cooling zone set If empty, keep the task pending deployment. The queued status is not used to issue deployment instructions, and the candidate conditions are re-examined in the next unified cycle.

[0091] If the candidate cooling zone set If not empty, it is a set of candidate cooling partitions. Set delay weights for each candidate cooling partition:

[0092] ;

[0093] In the formula, Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat conduction delay period; Represents the delay matrix The maximum value of the effective heat conduction delay period; Indicates the affected cooling zone number; This indicates the candidate cooling zone number. (For candidate cooling zones...) Tasks to be deployed The heat generated passes through After a unified cycle, it begins to affect the cooling zones. The power of the cooling equipment The smaller the value, the earlier the effect appears during task execution. Therefore, this invention sets a larger delay weight for the corresponding heat effect transfer coefficient, so that the smaller the heat conduction delay period, the larger the delay weight, and the delay weight corresponding to the effective heat effect is always greater than 0. When all effective heat conduction delay periods are 0, the delay matrix... Maximum value of effective heat conduction delay period The value is 0, at which point the candidate cooling zone set is... The delay weight of each candidate cooling partition is 1.

[0094] S6: Determine the optimal deployment location based on the expected cooling energy consumption of each candidate cooling zone and generate the initial target power value of the cooling equipment for each cooling zone.

[0095] It should be noted that the task to be deployed Deploy to candidate cooling partitions Subsequently, the expected maximum IT operating power consumption of the task. It will pass through candidate cooling zones This invention calculates the thermal impact on each cooling zone of the entire computer room, therefore, it calculates the thermal impact on each task to be deployed separately. The expected cooling energy consumption of the task when deployed to each candidate cooling partition is determined, and the candidate cooling partition with the lowest expected cooling energy consumption is determined as the optimal deployment location.

[0096] Specifically, the tasks to be deployed are determined based on the expected cooling energy consumption of the tasks corresponding to each candidate cooling zone. Optimal deployment location:

[0097] ;

[0098] In the formula, Indicates tasks to be deployed The optimal deployment location; Indicates tasks to be deployed The set of candidate cooling partitions; Indicates tasks to be deployed Estimated runtime, estimated task runtime Used to convert incremental electrical power of cooling equipment into expected cooling energy consumption during mission operation; Indicates tasks to be deployed The expected maximum IT operating power consumption; Indicates the total number of cooling zones; Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat transfer coefficient; This indicates a candidate cooling zone that minimizes the calculation result within the parentheses. To deploy tasks Deploy to candidate cooling partitions The expected cooling energy consumption during the mission.

[0099] Tasks to be deployed Deploy to the optimal deployment location The corresponding task's expected cooling energy consumption Recorded as .

[0100] Determine the optimal deployment location Then, for each cooling zone in the entire computer room Tasks to be deployed Expected maximum IT operating power consumption Multiply by the optimal deployment location For cooling zones Delay weight and heat transfer coefficient To obtain cooling zones The initial power increment of the cooling equipment. For optimal deployment location. If there is no effective thermal influence between cooling zones, and the thermal influence transfer coefficient is 0, then the corresponding initial power increment of the cooling equipment is 0.

[0101] Within the unified period when the optimal deployment location is determined, the cooling zone... The average power of the cooling equipment is used as the reference value for the power of the cooling equipment in this cooling zone. This reference value is then added to the corresponding initial power increment of the cooling equipment to obtain the cooling zone. The initial target power of the cooling equipment is obtained. Similarly, the initial target power of the cooling equipment in each cooling zone of the entire computer room is obtained.

[0102] Output optimal deployment location Optimal deployment location Expected cooling energy consumption for the mission And the initial target power values ​​for the cooling equipment in each cooling zone.

[0103] S7: Executes the deployment of tasks to be deployed and the cooling control of each cooling zone, and performs closed-loop correction based on the prediction error of the power change of the cooling equipment and the deviation of the air intake temperature.

[0104] Specifically, the optimal deployment location The task scheduler sends the initial target power values ​​for the cooling equipment in each cooling zone to the corresponding cooling system controller. Assign to optimal deployment location Each cooling system controller adjusts the electrical power of the cooling equipment according to the corresponding target value.

[0105] Within each unified cycle, the actual inlet air temperature and actual cooling equipment power of each cooling zone are collected. For each affected cooling zone... The actual change in cooling equipment power during this unified period is subtracted from the predicted change in cooling equipment power calculated according to step S3 to obtain the corresponding prediction error of cooling equipment power change. When the heat-affected zone... The absolute value of the IT power consumption change in at least one cooling zone exceeds the corresponding power consumption change dead zone threshold, and the cooling zone When all valid data markers for the cooling zones involved in the calculation are valid, the spatial heat transfer coefficient matrix is ​​updated using the recursive least squares method in step S3. The corresponding heat transfer coefficients are determined, and the delay matrix is ​​updated according to the error comparison method in step S3. The corresponding heat conduction delay period is determined; if the above conditions are not met, the corresponding heat influence transfer coefficient and heat conduction delay period remain unchanged.

[0106] When the target power value of the cooling equipment is adjusted due to inlet air temperature control, the unified period in which the target value is adjusted and the subsequent unified period during which the cooling equipment performs the adjustment are marked as the control adjustment period. The changes in the cooling equipment's power value within the control adjustment period are mainly caused by temperature control commands, and the spatial heat transfer coefficient matrix is ​​not updated using the corresponding changes in the cooling equipment's power value. and delay matrix When the change in the cooling medium flow rate does not exceed the minimum effective detectable change of the flow meter for two consecutive unified cycles, the cooling equipment is deemed to have completed this adjustment, and parameter updates will resume from the next unified cycle.

[0107] The difference between the average actual inlet air temperature of each cooling zone and the target inlet air temperature within the same unified cycle is taken as the inlet air temperature deviation. This inlet air temperature deviation is only used to increase, decrease, or maintain the target power value of the cooling equipment in the corresponding cooling zone according to the temperature control dead zone, and does not participate in the spatial heat transfer coefficient matrix. and delay matrix Parameter updates.

[0108] When the actual inlet air temperature exceeds the preset upper limit of the temperature control dead zone, the target power value of the cooling equipment in the corresponding cooling zone will be increased by 2% of the sum of the rated power of all cooling equipment in that cooling zone in the next unified cycle. When the actual inlet air temperature is below the preset lower limit of the temperature control dead zone, the corresponding target value will be decreased by 2% of the sum of the rated power of all cooling equipment in that cooling zone. When the actual inlet air temperature is within the preset temperature control dead zone, the corresponding target value will remain unchanged. The adjusted target power value of the cooling equipment must not be lower than 0 and must not be higher than the sum of the rated power of all cooling equipment in the corresponding cooling zone. In this embodiment, the target inlet air temperature is set to 24 degrees Celsius, and the temperature control dead zone is set to... 0.5 degrees Celsius, meaning the upper limit of the temperature control dead zone is 24.5 degrees Celsius and the lower limit is 23.5 degrees Celsius. In other embodiments, the target inlet air temperature can be set according to the server's allowable inlet air temperature range and operational requirements, and the temperature control dead zone and single-cycle adjustment ratio can be set according to the temperature sensor accuracy, cooling equipment response time, and allowable temperature fluctuations.

[0109] Running tasks maintain their current deployment location. When a new task is to be deployed, steps S5 and S6 are executed. After a running task ends, the operating status of the cooling equipment changes, or the available heat removal capability of the cooling partition changes, the candidate cooling partition set is updated, and steps S5 and S6 are re-executed for the queued tasks to be deployed. When the task scheduler supports task migration, if the cooling equipment in the cooling partition where the running task is located enters a fault or maintenance state, the running task is treated as a task to be deployed and steps S5 and S6 are re-executed.

[0110] In each uniform cycle, data acquisition, data validity judgment, calculation of cooling equipment power prediction error, parameter updating, and temperature control are repeatedly performed. As the operating time accumulates, the spatial heat transfer coefficient matrix... and delay matrix It continuously approximates the actual physical coupling characteristics under the current operating conditions. When the operating conditions change again, the new deviation drives a new round of parameter updates, enabling the system to have long-term adaptive tracking capabilities for time-varying environments, ultimately achieving intelligent energy-saving scheduling.

[0111] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise explicitly specified.

[0112] While this specification has shown and described numerous embodiments of the invention, 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. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for monitoring energy consumption and intelligent energy-saving scheduling in a computing center, characterized in that, include: Obtain operational status data for each cooling zone within the computing center, as well as the expected maximum IT power consumption and estimated runtime of tasks to be deployed; Based on the operating status data of each cooling zone, the heat influence transfer relationship and heat conduction delay relationship between each cooling zone are identified, and the spatial heat influence transfer coefficient matrix and delay matrix are constructed. Based on the spatial thermal effect transfer coefficient matrix, delay matrix, and communication topology requirements of the task to be deployed, a set of candidate cooling partitions that meet the deployment conditions is selected, and the delay weight of each candidate cooling partition to other cooling partitions is determined. Combining latency weights, spatial thermal impact transfer coefficient matrix, and expected maximum IT operating power consumption, the expected cooling energy consumption of the entire data center is calculated when the task to be deployed is deployed to each candidate cooling zone, and the candidate cooling zone with the minimum expected cooling energy consumption is determined as the optimal deployment location. Based on the expected increase in cooling energy consumption corresponding to the optimal deployment location, update the target power value of the cooling equipment in each cooling zone to schedule the tasks to be deployed and control the cooling equipment.

2. The method for monitoring and intelligent energy-saving scheduling of computing center energy consumption according to claim 1, characterized in that, The acquisition of operational status data for each cooling zone within the computing center includes: Real-time power consumption, temperature, cooling medium flow rate, and cooling equipment power of each cooling zone are collected; the collected raw data is processed by periodic segmentation, timestamp alignment, outlier removal, and missing value completion to generate a standard time series with valid data labels.

3. The method for monitoring and intelligent energy-saving scheduling of computing center energy consumption according to claim 2, characterized in that, The process of identifying the heat transfer relationships and heat conduction delay relationships between cooling zones based on their operational status data includes: Based on standard time series, the power consumption change and cooling equipment power change of each cooling zone are obtained; the thermal influence neighborhood of each cooling zone is determined according to the cold aisle layout, hot aisle layout and airflow organization direction of the computer room; using a recursive least squares algorithm with forgetting factor and regularization term, the thermal influence transfer coefficient and heat conduction delay period of each source cooling zone to the affected cooling zone are identified according to the power consumption change and cooling equipment power change of each cooling zone in the thermal influence neighborhood.

4. The method for monitoring and intelligent energy-saving scheduling of computing center energy consumption according to claim 3, characterized in that, The power consumption change of each cooling zone within the heat-affected zone and the power consumption change of the cooling equipment satisfy the following relationship: ； In the formula, Indicates period Internal cooling partition The change in electrical power of the cooling equipment; Cooling partition For cooling zones The heat conduction delay period; Cooling partition For cooling zones The heat transfer coefficient; Indicates the affected cooling zone number; Indicates the cooling zone number of the affected source; Indicates the current cycle number; Indicates the heat-affected neighborhood; Indicates period Internal cooling partition The change in IT power consumption.

5. The method for monitoring and intelligent energy-saving scheduling of computing center energy consumption according to claim 1, characterized in that, The construction of the space thermal effect transmission coefficient matrix and delay matrix includes: Arrange the heat transfer coefficients in the order that rows represent affected cooling zones and columns represent affected cooling zones to generate a spatial heat transfer coefficient matrix; arrange the heat conduction delay periods in the order that rows represent affected cooling zones and columns represent affected cooling zones to generate a delay matrix.

6. The method for monitoring and intelligent energy-saving scheduling of computing center energy consumption according to claim 1, characterized in that, The process involves selecting a set of candidate cooling partitions that meet the deployment conditions based on the spatial thermal effect transfer coefficient matrix, the delay matrix, and the communication topology requirements of the task to be deployed, and determining the delay weight of each candidate cooling partition relative to other cooling partitions, including: Cooling zones that simultaneously meet the network communication hop limit, normal operating status of cooling equipment, and sufficient heat removal capability are added to the candidate cooling zone set. For each candidate cooling zone in the candidate cooling zone set, the delay weight of the candidate cooling zone to other cooling zones is calculated based on the heat conduction delay period and its maximum value in the delay matrix.

7. The method for monitoring energy consumption and intelligent energy-saving scheduling of a computing center according to claim 6, characterized in that, The delay weight satisfies the expression: ； In the formula, Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat conduction delay period; This represents the maximum effective heat conduction delay period in the delay matrix; Indicates the affected cooling zone number; This indicates the candidate cooling zone number.

8. The method for monitoring energy consumption and intelligent energy-saving scheduling of computing centers according to claim 1, characterized in that, The optimal deployment location satisfies the expression: ； In the formula, Indicates tasks to be deployed The optimal deployment location; Indicates tasks to be deployed The set of candidate cooling partitions; Indicates tasks to be deployed The estimated runtime; Indicates tasks to be deployed The expected maximum IT operating power consumption; Indicates the total number of cooling zones; Indicates candidate cooling zones For cooling zones Delay weights; Indicates candidate cooling zones For cooling zones The heat transfer coefficient; This indicates a candidate cooling zone that minimizes the calculation result within the parentheses. To deploy tasks Deploy to candidate cooling partitions The expected cooling energy consumption during the mission.

9. The method for monitoring energy consumption and intelligent energy-saving scheduling of computing centers according to claim 1, characterized in that, The step of updating the target power values ​​of the cooling equipment in each cooling zone based on the expected cooling energy consumption increment corresponding to the optimal deployment location includes: For each cooling zone, the expected maximum IT operating power consumption is multiplied by the delay weight and heat transfer coefficient of the optimal deployment location for that cooling zone to obtain the initial cooling equipment power increment; The current baseline value of the cooling equipment's electrical power is added to the initial increment of the cooling equipment's electrical power to obtain the target value of the cooling equipment's electrical power for the corresponding cooling zone.

10. The method for monitoring energy consumption and intelligent energy-saving scheduling of computing centers according to claim 1, characterized in that, After scheduling tasks to be deployed and controlling the cooling equipment, the process also includes: The actual power variation of the cooling equipment in each cooling zone is collected, and the prediction error of the power variation is calculated. Based on the prediction error of the power variation, the parameters in the spatial heat influence transfer coefficient matrix and the delay matrix are updated online. The inlet air temperature deviation of each cooling zone is obtained, and the target power of the cooling equipment in each cooling zone is dynamically adjusted based on the inlet air temperature deviation.

Images