Intelligent scheduling method and system for IDC machine room energy efficiency optimization

By constructing a cooling margin time-series slice matrix and task hot-launch profiles, complementary task pairs that are executed concurrently are identified, and a computing-cooling collaborative scheduling scheme for IDC data centers is generated. This solves the problem of time differences between computing resources and cooling systems, and improves energy efficiency and thermal security.

CN123068086APending Publication Date: 2026-11-10北京航天尚讯科技有限公司
0 Cites 0 Cited by

Patent Information

Application Number
CN202611232385.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

The time scale difference between computing resource scheduling and cooling system in existing IDC data centers leads to the inability to fully utilize cooling margins, and the failure to effectively coordinate and optimize task hot launch characteristics and cooling response results in low energy efficiency and hot spot risk.

Method used

By constructing a cooling margin time-series slice matrix and a task hot-launch time-series profile, complementary task pairs are identified for concurrent execution. Combined with global minimum energy consumption and thermal safety constraints, a computation-cooling collaborative scheduling scheme is generated.

Benefits of technology

This achieves efficient utilization of the cooling system, reduces energy consumption and operating costs, enhances the system's robustness to dynamic loads and temperature fluctuations, and ensures thermal safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068086A_ABST
    Figure CN123068086A_ABST
Patent Text Reader

Abstract

The present application relates to IDC machine room energy efficiency optimization technical field, especially relate to a kind of IDC machine room energy efficiency optimization intelligent scheduling method and system.The cooling residual matrix time slice matrix is constructed by the cooling partition working condition periodic change and ambient temperature pre-computing available cooling residual, the heat emission time profile of the task to be scheduled is extracted according to calculation stage disassembly and two two phase analysis is carried out to obtain paired superposition heat emission profile, the space-time alignment of paired profile and unpaired profile with cooling residual matrix is carried out, and the space-time placement position that satisfies envelope coverage is found, then the global energy consumption is lowest and heat safety is taken as target joint optimization calculation resource allocation and cooling capacity response, generates IDC machine room calculation-cooling collaborative scheduling scheme;The present application can reduce global energy consumption and meet the heat safety constraint, improve IDC machine room energy efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of energy efficiency optimization technology for IDC data centers, and in particular to an intelligent scheduling method and system for energy efficiency optimization of IDC data centers. Background Technology

[0002] Current energy efficiency optimization in data centers typically employs a strategy of separating compute resource scheduling from cooling system management. In conventional practices, the scheduling center predicts future computing demand based on real-time load or historical data and allocates tasks to different servers or racks, while the cooling system independently adjusts the output of cooling equipment based on feedback from rack inlet temperature sensors. Some solutions incorporate thermodynamic models to dynamically allocate cooling capacity based on temperature field distribution; however, a time scale difference still exists between compute scheduling and cooling response—compute load changes occur on the order of seconds or even milliseconds, while the cooling system response lags by several minutes. This separation model results in the underutilization of cooling capacity, with a large amount of cooling energy wasted during task idle periods, while facing the risk of localized hotspots during peak computing periods.

[0003] Another common approach is to treat task scheduling and cooling optimization as two independent problems, solving them separately with the goal of minimizing energy consumption. For example, first reducing the load on compute nodes through virtual machine migration or task merging, and then independently optimizing cooling system parameters. This approach ignores the complementarity between task hot-launch characteristics and cooling headroom over time. In actual operation, the heat intensity varies significantly across different computational phases (such as data loading, intensive computation, and I / O wait), but existing scheduling mechanisms only focus on the overall computational load or average power consumption of a task, without performing internal phase decomposition and hot-launch timing analysis. This leads to the cooling system over-responding during brief periods of high heat generation, while cooling resources remain idle during low-heat generation, resulting in overall energy efficiency falling short of optimal levels.

[0004] Current technologies lack the ability to collaboratively optimize the thermal emission timing of multiple tasks. When multiple tasks execute concurrently within the same cooling zone, overlapping peaks in their thermal emission curves can drastically increase local cooling pressure; conversely, complementary distributions can reduce peak demand. However, conventional scheduling does not perform phase matching analysis on task pairs, nor can it utilize this temporal complementarity to smooth out thermal load fluctuations, thus limiting the potential of the cooling system to utilize natural cooling sources or efficient operating modes. Summary of the Invention

[0005] This invention provides an intelligent scheduling method and system for optimizing energy efficiency in IDC (Internet Data Center) data centers, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an intelligent scheduling method for optimizing energy efficiency in IDC (Internet Data Center) data centers, comprising:

[0007] Based on the periodic change pattern of equipment operating conditions in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin of each cooling zone is pre-calculated on a time-by-time basis, and a time-series slice matrix of cooling margin is constructed.

[0008] The task to be scheduled is broken down into internal calculation stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled.

[0009] Pairwise phase analysis is performed on the thermal emission timing profiles of each task to be scheduled in the task queue to identify two pairs of tasks to be scheduled that have complementary thermal emission distributions within the same candidate time window and arrange them to be executed concurrently on adjacent computing nodes in the same cooling zone. The thermal emission intensities of the two tasks to be scheduled on each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles.

[0010] For tasks that have been paired and scheduled, take the paired superimposed thermal emission profile; for tasks that have not been paired and scheduled, take the thermal emission timing profile. Spatiotemporally align these profiles with the cooling margin timing slice matrix. Slide candidate starting slices along the time axis to find the spatiotemporal placement position where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension.

[0011] Starting from the aforementioned spatiotemporal placement location, and with the goal of minimizing global energy consumption and meeting thermal safety constraints, a collaborative scheduling scheme for computing and cooling capacity in the IDC data center is generated by jointly optimizing the allocation of computing resources and the response of cooling capacity.

[0012] Based on the periodic variation patterns of equipment operating conditions in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin for each cooling zone is pre-calculated on a time-series basis, and a time-series slice matrix of cooling margin is constructed, including:

[0013] The historical operating data of the equipment in each cooling zone of the IDC data center is periodically decomposed, and the periodic reference component and the deviation component of the current time period of each equipment operating parameter are extracted. The periodic reference component is superimposed on the deviation component of the current time period to obtain the periodic change pattern of the equipment operating conditions in each cooling zone.

[0014] Based on the periodic variation pattern and the measured ambient temperature at the current moment, the predicted ambient temperature value corresponding to each time slice is calculated backward;

[0015] Based on the positional relationship of time slices, the predicted ambient temperature values ​​of adjacent time slices and non-adjacent time slices are obtained. The upper limit of the available cooling margin of adjacent time slices is corrected by the predicted ambient temperature values ​​of adjacent time slices, and the upper limit of the available cooling margin of non-adjacent time slices is constrained by the predicted ambient temperature values ​​of non-adjacent time slices, thus obtaining the graded corrected upper limit of the available cooling margin for each time slice.

[0016] Based on the periodic variation pattern and the upper bound of the available cooling margin for graded correction, the available cooling margin of each cooling zone is pre-calculated on a time-by-time slice basis, and the time-series slice matrix of the cooling margin is constructed.

[0017] The task to be scheduled is broken down into internal computational stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled, including:

[0018] The resource usage characteristics at each moment in the historical execution record of the task to be scheduled are analyzed. The abrupt change position of the resource usage characteristics on the time axis is taken as the boundary of the internal calculation stage. The task to be scheduled is decomposed into multiple sequentially arranged internal calculation stages to obtain the stage boundary sequence of each internal calculation stage.

[0019] The resource occupancy characteristic intervals of each internal computing stage are defined by the stage boundary sequence, and the resource occupancy characteristics within each internal computing stage interval are aggregated to obtain the resource occupancy benchmark characteristics of each internal computing stage.

[0020] Using the resource occupancy benchmark features of each internal calculation stage in the historical execution record of the task to be scheduled and the corresponding historical thermal emission record values ​​as samples, a mapping relationship between the resource occupancy benchmark features and thermal emission intensity is established;

[0021] By taking the resource consumption baseline characteristics of each internal computing stage of the currently scheduled task as input and substituting them into the established mapping relationship, the thermal emission intensity of each internal computing stage can be obtained.

[0022] The duration of each internal computation stage is calculated based on the start and end boundaries of each internal computation stage in the stage boundary sequence. The thermal emission intensity and duration of each internal computation stage are combined in stage order to obtain the thermal emission timing profile of the task to be scheduled.

[0023] Pairwise phase analysis is performed on the thermal emission timing profiles of each scheduled task in the task queue to identify two pairs of scheduled tasks with complementary thermal emission distributions within the same candidate time window. These pairs are then arranged to be executed concurrently on adjacent computing nodes within the same cooling zone. The thermal emission intensities of the two scheduled tasks in each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles, including:

[0024] Extract the extreme positions and extreme amplitudes of the thermal emission intensity from each time slice to obtain the thermal emission phase feature sequence of each task to be scheduled;

[0025] Using the candidate time window as the alignment benchmark, the thermal emission phase feature sequences of each task to be scheduled in the task queue are compared pairwise. The degree of offset and complementarity of the extreme position of the two tasks to be scheduled are calculated in time slices. The degree of offset and complementarity in time slices are combined into a thermal emission complementarity metric value of the task pair in the candidate time window.

[0026] Using the thermal emission complementarity metric of each task pair as a criterion, task pairs whose thermal emission complementarity metric values ​​satisfy the complementary distribution condition within the same candidate time window are identified, thus obtaining a set of thermal emission complementary task pairs.

[0027] Each thermal emission complementary task pair in the set of thermal emission complementary task pairs is arranged to be executed concurrently on adjacent computing nodes within the same cooling partition. According to the time slice order within the candidate time window, the thermal emission intensity of the two tasks to be scheduled in each thermal emission complementary task pair is added slice by slice to obtain the paired superimposed thermal emission profile of each thermal emission complementary task pair.

[0028] For scheduled tasks that have completed pairing, the paired superimposed thermal emission profile is taken; for scheduled tasks that have not completed pairing, the thermal emission timing profile is taken. Spatiotemporal alignment with the cooling margin timing slice matrix is ​​performed, including:

[0029] Each task to be scheduled in the task queue is distinguished according to its pairing status. The paired superimposed thermal emission profile is extracted as the placement evaluation profile for the task to be scheduled that has been paired. The thermal emission timing profile is extracted as the placement evaluation profile for the task to be scheduled that has not been paired, thus obtaining a set of placement evaluation profiles for each task to be scheduled.

[0030] Using the cooling partition affiliation corresponding to the placement evaluation contour of each scheduled task as an index, the available cooling margin sequence of the relevant cooling partition dimension corresponding to each scheduled task is extracted from the cooling margin time-series slice matrix to obtain the set of available cooling margin sequences of each scheduled task's cooling partition.

[0031] Using the time slice length of the placement evaluation profile of each task to be scheduled as the alignment window, the placement evaluation profile of each task to be scheduled in the placement evaluation profile set is aligned with the corresponding available cooling margin sequence in the available cooling margin sequence set of the cooling partition at the time slice granularity, so as to obtain the spatiotemporal alignment result of the placement evaluation profile of each task to be scheduled and the corresponding available cooling margin sequence.

[0032] Slide candidate starting slices along the time axis to find spatiotemporal placement positions where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension, including:

[0033] Based on the spatiotemporal alignment results, the starting point of the time axis of the cooling margin time-series slice matrix is ​​taken as the initial candidate starting slice, and the candidate placement interval is defined on the time axis of the cooling margin time-series slice matrix according to the time slice length of the placement evaluation profile of each task to be scheduled.

[0034] The candidate starting slice is moved step by step along the time axis. For each candidate starting slice and the candidate placement interval corresponding to the candidate placement interval, the available cooling margin of each relevant cooling partition dimension in the cooling margin time slice matrix is ​​extracted on the corresponding time slice. The heat emission intensity of the placement evaluation contour on the corresponding time slice is compared on each relevant cooling partition dimension to obtain the time slice and cooling partition dimension envelope coverage determination result corresponding to each candidate starting slice.

[0035] The effective placement condition is that all time-series slice envelope coverage determination results in each relevant cooling partition dimension meet the requirement that the available cooling margin covers the thermal emission intensity of the placement evaluation contour. Candidate starting slices that meet the effective placement condition are extracted from the envelope coverage determination results of all candidate starting slices. The candidate starting slices that meet the effective placement condition are combined with the corresponding cooling partition affiliation information to obtain the spatiotemporal placement position of each task to be scheduled.

[0036] Starting from the aforementioned spatiotemporal placement location, and aiming at minimizing global energy consumption and meeting thermal safety constraints, a collaborative scheduling scheme for IDC data center computing and cooling capacity is generated by jointly optimizing computing resource allocation and cooling capacity response, including:

[0037] Starting from the spatiotemporal placement location, the computational load intensity distribution and cooling demand intensity distribution of each task to be scheduled at the spatiotemporal placement location are extracted. With the minimum global energy consumption as the optimization objective and the thermal safety constraint as the feasible domain boundary, the computational resource allocation dimension and the cooling capacity response dimension are incorporated into the same joint optimization framework to construct a joint optimization objective covering the computational resource allocation dimension and the cooling capacity response dimension.

[0038] Under the constraint of the joint optimization objective, the computational load intensity distribution of each task to be scheduled in the computational resource allocation dimension is quantified to obtain the computational resource allocation scheme of each task to be scheduled; taking the heat load change of each cooling zone corresponding to the computational resource allocation scheme as input, the cooling capacity response of each cooling zone in the cooling capacity response dimension is collaboratively optimized to obtain a cooling capacity response scheme that matches the computational resource allocation scheme.

[0039] Using the convergence criterion that both the computing resource allocation scheme and the cooling capacity response scheme simultaneously satisfy thermal safety constraints and minimize global energy consumption, the computing resource allocation scheme and the cooling capacity response scheme are iteratively verified to obtain the converged computing resource allocation result and cooling capacity response result.

[0040] The converged computing resource allocation results and cooling capacity response results are integrated with the spatiotemporal placement location to generate an IDC data center computing-cooling collaborative scheduling scheme.

[0041] A second aspect of this invention provides an intelligent scheduling system for optimizing energy efficiency in IDC (Internet Data Center) data centers, comprising:

[0042] The cooling margin unit is used to pre-calculate the available cooling margin of each cooling zone based on the periodic change pattern of the operating conditions of the equipment in each cooling zone of the IDC data center and the predicted value of the ambient temperature, and to construct a cooling margin time-series slice matrix.

[0043] The hot emission wheel unit is used to decompose the task to be scheduled according to the internal calculation stage, extract the hot emission intensity and duration of each stage, and obtain the hot emission timing profile of the task to be scheduled.

[0044] The pairing and overlay unit is used to perform pairwise phase analysis on the thermal emission timing profiles of each task to be scheduled in the task queue, identify two pairs of tasks to be scheduled whose thermal emission is complementary in the same candidate time window, and arrange them to be executed concurrently on adjacent computing nodes in the same cooling partition. The thermal emission intensity of the two tasks to be scheduled on each time slice is added slice by slice to obtain the pairing and overlay thermal emission profiles.

[0045] The spatiotemporal alignment unit is used to take the paired superimposed thermal emission profile of the scheduled tasks that have been paired and the thermal emission timing profile of the scheduled tasks that have not been paired, and perform spatiotemporal alignment with the cooling margin timing slice matrix respectively. The candidate starting slice is slid along the time axis to find the spatiotemporal placement position that makes the paired superimposed thermal emission profile or the thermal emission timing profile covered by the available cooling margin envelope in each relevant cooling partition dimension.

[0046] The collaborative scheduling unit is used to generate an IDC data center computing-cooling collaborative scheduling scheme by taking the spatiotemporal placement location as the starting point, aiming at minimizing global energy consumption and meeting thermal safety constraints, and jointly optimizing the allocation of computing resources and cooling capacity response.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] The beneficial effects of this application are as follows:

[0053] Based on the periodic changes in the operating conditions of equipment in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling capacity is pre-calculated time-by-time slices, and a time-series matrix is ​​constructed. This allows the scheduling process to proactively utilize the idle capacity of the cooling system, avoiding over-allocation or idleness of cooling resources. Tasks are broken down by computational stage, and the heat emission intensity and duration are extracted to form a precise heat emission time-series profile. This lays the foundation for fine-tuning heat distribution and cooling capacity, thereby suppressing the generation of local hotspots at the source and reducing unnecessary cooling energy consumption.

[0054] By pairing two scheduled tasks with complementary thermal emission distributions within the candidate time window and executing them concurrently on adjacent computing nodes in the same cooling partition, the thermal emissions of the two tasks cancel each other out or stagger peaks on the same time slice, significantly reducing the joint thermal emission peak. The paired superimposed thermal emission profile is spatiotemporally aligned with the cooling margin time series matrix, and the sliding candidate starting slice ensures that the thermal emission profile is covered by the available cooling margin in each relevant cooling partition dimension. This fully utilizes the temporal and spatial margins of the cooling system, achieving a dynamic balance between heat load and cooling capacity, and reducing frequent start-stop and over-adjustment energy consumption of the cooling system caused by thermal shock.

[0055] With the goals of minimizing global energy consumption and meeting thermal safety constraints, this approach jointly optimizes computing resource allocation and cooling capacity response to generate a compute-cooling collaborative scheduling scheme. While meeting equipment temperature limits, this scheme proactively schedules computing tasks to spatiotemporal regions with ample cooling capacity, enabling the cooling system to provide the necessary cooling with lower power and more stable operation. This avoids the energy waste caused by the disconnect between computing load and cooling response in traditional independent scheduling. The overall approach significantly improves the power utilization efficiency (PUE) of data centers, reduces operating costs, and enhances the system's robustness to dynamic loads and temperature fluctuations, ensuring thermal safety boundaries under high-density deployments. Attached Figure Description

[0056] Figure 1 A flowchart illustrating an intelligent scheduling method for optimizing energy efficiency in IDC data centers;

[0057] Figure 2A flowchart illustrating the process of generating a collaborative scheduling scheme for computing and cooling in an IDC (Internet Data Center) server room. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0059] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0060] Figure 1 This is a flowchart illustrating the intelligent scheduling method for optimizing energy efficiency in IDC data centers according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Based on the periodic change pattern of equipment operating conditions in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin of each cooling zone is pre-calculated on a time-by-time basis, and a time-series slice matrix of cooling margin is constructed.

[0062] The task to be scheduled is broken down into internal calculation stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled.

[0063] Pairwise phase analysis is performed on the thermal emission timing profiles of each task to be scheduled in the task queue to identify two pairs of tasks to be scheduled that have complementary thermal emission distributions within the same candidate time window and arrange them to be executed concurrently on adjacent computing nodes in the same cooling zone. The thermal emission intensities of the two tasks to be scheduled on each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles.

[0064] For tasks that have been paired and scheduled, take the paired superimposed thermal emission profile; for tasks that have not been paired and scheduled, take the thermal emission timing profile. Spatiotemporally align these profiles with the cooling margin timing slice matrix. Slide candidate starting slices along the time axis to find the spatiotemporal placement position where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension.

[0065] Starting from the aforementioned spatiotemporal placement location, and with the goal of minimizing global energy consumption and meeting thermal safety constraints, a collaborative scheduling scheme for computing and cooling capacity in the IDC data center is generated by jointly optimizing the allocation of computing resources and the response of cooling capacity.

[0066] In one optional implementation, based on the periodic variation pattern of the operating conditions of equipment in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin of each cooling zone is pre-calculated on a time-by-time basis, and a cooling margin time-series slice matrix is ​​constructed, including:

[0067] The historical operating data of the equipment in each cooling zone of the IDC data center is periodically decomposed, and the periodic reference component and the deviation component of the current time period of each equipment operating parameter are extracted. The periodic reference component is superimposed on the deviation component of the current time period to obtain the periodic change pattern of the equipment operating conditions in each cooling zone.

[0068] Based on the periodic variation pattern and the measured ambient temperature at the current moment, the predicted ambient temperature value corresponding to each time slice is calculated backward;

[0069] Based on the positional relationship of time slices, the predicted ambient temperature values ​​of adjacent time slices and non-adjacent time slices are obtained. The upper limit of the available cooling margin of adjacent time slices is corrected by the predicted ambient temperature values ​​of adjacent time slices, and the upper limit of the available cooling margin of non-adjacent time slices is constrained by the predicted ambient temperature values ​​of non-adjacent time slices, thus obtaining the graded corrected upper limit of the available cooling margin for each time slice.

[0070] Based on the periodic variation pattern and the upper bound of the available cooling margin for graded correction, the available cooling margin of each cooling zone is pre-calculated on a time-by-time slice basis, and the time-series slice matrix of the cooling margin is constructed.

[0071] For each cooling zone in the IDC data center, operational data from the past several weeks was collected for equipment such as chiller units, chilled water pumps, precision air conditioners, and cooling towers, including multi-dimensional parameters such as supply and return water temperatures, fan speeds, compressor power, and cooling capacity. A time series decomposition method was used to break down the historical sequence of each equipment operating parameter into a periodic baseline component and a current-period deviation component. The periodic baseline component reflects the regular fluctuations of the equipment within a typical day or week, such as the stable operation of chiller units during low-load periods at night and their full-load operation during peak days. The current-period deviation component captures the deviation between the actual operating value and the historical baseline caused by recent equipment aging, localized faults, temporary maintenance, or abnormal loads. By overlaying the periodic baseline component and the current-period deviation component on a time-slice basis, the periodic variation pattern of the equipment operating conditions in each cooling zone was obtained. This pattern retains the long-term periodic characteristics of equipment operation while incorporating short-term measured deviation information, making subsequent pre-calculations closer to actual operating conditions.

[0072] Based on the obtained periodic variation pattern and the measured ambient temperature collected at the current moment, the predicted ambient temperature values ​​for each time slice are extrapolated. During the extrapolation process, the measured ambient temperature at the current moment is used as the initial anchor point, and the temperature change trend extracted from the periodic variation pattern is combined to gradually extend forward according to the time slice step size. Let the slice index at the current moment be... , No. The predicted ambient temperature value for each time slice is ,but Consistent with periodic reference temperature components Compared with the attenuated current deviation component Joint decision, that is ,in The deviation attenuation coefficient satisfies This indicates that as the distance between the predicted time slice and the current time increases, the impact of measured deviations on future predictions gradually weakens, and the predicted values ​​converge towards the periodic baseline component. This design makes the prediction of recent time slices more dependent on measured values, while the prediction of long-term time slices depends more on historical periodic patterns, which is consistent with the physical law that the uncertainty of temperature prediction increases with the increase of the prediction step size.

[0073] Based on their positional relationship with the current time, time slices are classified into two categories: adjacent time slices and non-adjacent time slices. Adjacent time slices are those that are close to the current time on the time axis and have high prediction confidence; their environmental temperature predictions have high reliability. Non-adjacent time slices are those that are far from the current time and have greater prediction uncertainty. For adjacent time slices, their environmental temperature predictions are used as the basis for classification. The upper limit of the available cooling margin for this slice is directly corrected by subtracting the efficiency loss on the condenser side due to the increase in ambient temperature and the additional available margin due to the decrease in ambient temperature from the rated cooling capacity, thus obtaining the temperature-corrected upper limit of the available cooling margin. For non-adjacent time slices, due to significant prediction uncertainty, a conservative constraint strategy is adopted, using the predicted ambient temperature value. A constraint is imposed on the upper limit of the available cooling margin, that is, an uncertainty margin factor is introduced on the basis of the rated cooling capacity. The upper limit of the available cooling margin for non-adjacent slices is compressed to... This ensures sufficient thermal safety margin even when prediction deviations are large. After the above-mentioned graded correction process, the upper bound of the available cooling margin for graded correction is obtained for each time slice. Adjacent slices use precise correction values, while non-adjacent slices use conservative constraint values. Together, they constitute the upper bound constraint input for subsequent slice-by-slice pre-calculation.

[0074] Based on the periodic variation pattern and the graded correction of the upper limit of available cooling margin, the available cooling margin of each cooling zone is pre-calculated on a time-slice basis. For each cooling zone... In the For each time slice, the expected operating status of the cooling equipment within that slice is first determined based on the periodic variation pattern, including the expected output cooling capacity of the refrigeration unit. The parameters include the expected air volume and temperature of the precision air conditioner, and the expected supply water temperature of the chilled water system; based on these, the cooling capacity already occupied by the scheduled tasks in the cooling zone within this slice is comprehensively considered. The available cooling margin of this partition in this slice is calculated. Its calculation expression is: ,in The upper limit of the cooling margin (taken from adjacent slices) can be used for the graded correction of the corresponding slices. Non-adjacent slices are taken The smaller of the two values ​​is taken to ensure that the actual available capacity is not exceeded. The actual available capacity is then obtained by subtracting the occupied capacity.

[0075] The available cooling capacity of each cooling zone at each time slice. A cooling capacity time-series slice matrix is ​​constructed, arranged by partition and time slice dimensions. Rows in the matrix correspond to each cooling partition, and columns correspond to each time slice. Matrix elements represent the available cooling capacity value of the corresponding partition within its corresponding slice. This matrix construction fully integrates the periodic patterns of equipment operating conditions, recent deviations in measured ambient temperature, and confidence intervals of cooling capacity at different time scales resulting from a tiered correction strategy. This provides a structured, multi-dimensional cooling capacity description foundation for the spatiotemporal alignment and matching of subsequent mission hot-launch timing profiles with cooling capacity. In real-world IDC data center scenarios, different cooling partitions are configured with different types of cooling equipment. For example, some partitions use air-cooled precision air conditioners, while others use liquid-cooled backplanes. Therefore, during partition-by-partition and slice-by-slice pre-calculation, corresponding cooling capacity calculation models must be established based on the specific equipment type of each partition. This ensures that the physical meaning of each element in the matrix matches the actual equipment capacity, avoiding overestimation or underestimation of cooling capacity due to differences in equipment type. This guarantees the thermal safety and energy efficiency optimization of subsequent scheduling decisions.

[0076] In one optional implementation, the task to be scheduled is decomposed into internal computational stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled, including:

[0077] The resource usage characteristics at each moment in the historical execution record of the task to be scheduled are analyzed. The abrupt change position of the resource usage characteristics on the time axis is taken as the boundary of the internal calculation stage. The task to be scheduled is decomposed into multiple sequentially arranged internal calculation stages to obtain the stage boundary sequence of each internal calculation stage.

[0078] The resource occupancy characteristic intervals of each internal computing stage are defined by the stage boundary sequence, and the resource occupancy characteristics within each internal computing stage interval are aggregated to obtain the resource occupancy benchmark characteristics of each internal computing stage.

[0079] Using the resource occupancy benchmark features of each internal calculation stage in the historical execution record of the task to be scheduled and the corresponding historical thermal emission record values ​​as samples, a mapping relationship between the resource occupancy benchmark features and thermal emission intensity is established;

[0080] By taking the resource consumption baseline characteristics of each internal computing stage of the currently scheduled task as input and substituting them into the established mapping relationship, the thermal emission intensity of each internal computing stage can be obtained.

[0081] The duration of each internal computation stage is calculated based on the start and end boundaries of each internal computation stage in the stage boundary sequence. The thermal emission intensity and duration of each internal computation stage are combined in stage order to obtain the thermal emission timing profile of the task to be scheduled.

[0082] Extracting the hot-launch timing profile of scheduled tasks hinges on transforming the differences in computational behavior within the task into a quantifiable description of hot-launch activity. Different types of computational tasks often undergo multiple significantly different computational phases during execution, such as data preprocessing, model forward inference, gradient backpropagation, and result writing. Each phase exhibits distinctly different characteristic distributions in resource dimensions such as CPU utilization, GPU utilization, memory bandwidth, and I / O throughput. Through in-depth analysis of historical execution records, these phase differences can be extracted in a structured manner, thereby establishing a mapping relationship from resource consumption characteristics to hot-launch intensity.

[0083] By analyzing historical execution records and sampling resource usage characteristics at each time point using equal-length windows, a multi-dimensional time series is formed. The characteristic variation amplitude between adjacent sampling points is calculated point-by-point along the time axis of this multi-dimensional time series. When the variation amplitude at a certain time point exceeds a preset threshold, that time point is marked as the boundary of the internal calculation stage. Specifically, let the resource usage feature vector be... ,in Here is the sampling time index, and the feature difference measure between adjacent sampling points is... ,when When the adaptive threshold is exceeded, These are recorded as boundary points. The adaptive threshold can be determined based on the statistical distribution of feature difference sequences in historical records, thereby avoiding false boundaries caused by noise. All boundary points are arranged in chronological order, and combined with the start and end times of the task to form a complete stage boundary sequence, thus decomposing the task to be scheduled into several sequentially arranged internal computation stages.

[0084] After obtaining the stage boundary sequence, the resource occupancy characteristics of all sampling times within the time interval defined by each boundary point are aggregated to obtain the resource occupancy baseline characteristics of each internal computation stage. The aggregation method can use the mean of the feature vectors of each sampling point within the interval, i.e., the mean of the feature vectors of the sampling points within the interval. Each internal computing stage has its own resource consumption baseline characteristics. The mean value of the feature vectors of all sampling points within this stage interval is obtained as follows: ,in For the first The set of sampling times corresponding to each stage This represents the number of sampling points in the set. For stages with a clear bimodal distribution, more robust aggregation strategies such as the median or weighted mean can be used to reduce the interference of outlier sampling points on the baseline features.

[0085] Establish a mapping relationship between resource occupancy baseline characteristics and thermal emission intensity, using resource occupancy baseline characteristics from each internal calculation stage in historical execution records. As input, and using historical thermal emission records of computing nodes within the corresponding stage (typically derived from measured data from embedded temperature sensors or power consumption monitoring modules within the nodes) as output, a regression mapping model is constructed. When the sample size is sufficient, nonlinear regression methods, such as gradient boosting trees or lightweight neural network-based regressors, can be used to capture the nonlinear relationship between resource occupancy characteristics and thermal emission intensity. When the sample size is limited, linear regression or ridge regression can be used as alternatives to ensure the stability of the mapping relationship. The mapping relationship is established using historical execution records as training samples, and the mapping accuracy is evaluated through cross-validation to ensure that the prediction error is within an acceptable range.

[0086] The resource consumption baseline characteristics of each internal computation stage of the currently scheduled task. Substituting the established mapping relationship, we obtain the predicted thermal emission intensity values ​​for each stage. The unit of this predicted value is consistent with the unit of heat load description in the IDC data center thermal management system, usually in kilowatts (kW), so as to facilitate direct comparison with the cooling margin time-series slice matrix. For new tasks with few historical execution records, the thermal emission intensity can be estimated by using migration mapping based on the task's resource consumption characteristics and the similarity with existing tasks, so as to ensure the coverage integrity of the thermal emission time-series profile.

[0087] Based on the start and end boundaries of each internal computational stage in the stage boundary sequence, calculate the duration of each stage. That is, the first The difference between the end boundary time and the start boundary time of each stage. Regarding time slice granularity alignment, the duration will be... Round up to the number of time slices This ensures that the thermal emission intensity of each stage can be described in units of integer time slices, maintaining consistency with the time granularity of the cooling margin time-series slice matrix. Within the same time slice, if a certain stage only occupies a portion of the slice's duration, the thermal emission intensity is weighted according to the proportion of that stage within the slice to improve the accuracy of the contour description.

[0088] The thermal emission intensity of each internal calculation stage With corresponding duration The phases are combined sequentially to form the thermal emission timing profile of the task to be scheduled. This profile uses time slices as the basic unit, with each slice corresponding to a thermal emission intensity value, forming a thermal emission intensity sequence unfolding along the time axis. For phases whose duration spans multiple time slices, the same thermal emission intensity value is filled into all slices covered by that phase. For stages whose duration is less than one time slice, they are superimposed on the corresponding slice using a weighted average. The resulting thermal emission timing profile fully describes the thermal emission intensity distribution across each time slice during the entire execution process of the scheduled task, providing accurate thermal load input for subsequent phase analysis and spatiotemporal placement.

[0089] In practical applications, when the same type of task is executed on different computing nodes, the intensity of thermal emission varies between nodes due to differences in node hardware (such as processor model, heat dissipation structure, memory configuration, etc.). To address this, node type identifiers can be introduced as an additional feature dimension when establishing the mapping relationship. This allows the mapping relationship to distinguish thermal emission differences under different node types, thereby improving the accuracy of the thermal emission timing profile in cross-node scheduling scenarios. For cases where task execution duration fluctuates significantly, historical statistical quantiles can be incorporated into the duration estimation as a conservative estimate to avoid underestimating the duration and causing actual thermal emission to exceed cooling margin constraints.

[0090] In one optional implementation, pairwise phase analysis is performed on the thermal emission timing profiles of each scheduled task in the task queue to identify two pairs of scheduled tasks with complementary thermal emission distributions within the same candidate time window. These pairs are then arranged to be executed concurrently on adjacent computing nodes within the same cooling partition. The thermal emission intensities of the two scheduled tasks on each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles, including:

[0091] Extract the extreme positions and extreme amplitudes of the thermal emission intensity from each time slice to obtain the thermal emission phase feature sequence of each task to be scheduled;

[0092] Using the candidate time window as the alignment benchmark, the thermal emission phase feature sequences of each task to be scheduled in the task queue are compared pairwise. The degree of offset and complementarity of the extreme position of the two tasks to be scheduled are calculated in time slices. The degree of offset and complementarity in time slices are combined into a thermal emission complementarity metric value of the task pair in the candidate time window.

[0093] Using the thermal emission complementarity metric of each task pair as a criterion, task pairs whose thermal emission complementarity metric values ​​satisfy the complementary distribution condition within the same candidate time window are identified, thus obtaining a set of thermal emission complementary task pairs.

[0094] Each thermal emission complementary task pair in the set of thermal emission complementary task pairs is arranged to be executed concurrently on adjacent computing nodes within the same cooling partition. According to the time slice order within the candidate time window, the thermal emission intensity of the two tasks to be scheduled in each thermal emission complementary task pair is added slice by slice to obtain the paired superimposed thermal emission profile of each thermal emission complementary task pair.

[0095] After obtaining the thermal emission timing profiles of each task to be scheduled, pairwise phase analysis needs to be performed on all tasks in the task queue to identify task pairs with complementary thermal emission characteristics in terms of temporal distribution. The core of phase analysis lies in characterizing the fluctuation pattern of thermal emission intensity for each task across time slices, rather than focusing solely on the overall average heating level. Therefore, for each task to be scheduled, its thermal emission timing profile is scanned slice by slice, extracting the extreme positions and amplitudes of thermal emission intensity in each time slice, forming a thermal emission phase feature sequence for that task. The extreme position describes the relative location of the thermal emission peak or trough on the time axis, while the extreme amplitude quantifies the magnitude of the heating intensity at that extreme point. Together, they constitute a compact description of the task's thermal emission waveform structure, providing a usable feature basis for subsequent pairwise comparisons.

[0096] Using candidate time windows as alignment benchmarks, a slice-by-slice comparison is performed on the hot-emission phase feature sequences of any two tasks to be scheduled in the task queue. Let the task... With the task Coverage within the candidate time window The time slice, for the first time slice slice Extract tasks separately extreme position With extreme value amplitude and tasks extreme position With extreme value amplitude Degree of offset Defined as the absolute value of the difference between the extreme locations of the two tasks on this slice, i.e. The larger the value, the greater the temporal stagger between the thermal emission peaks of the two tasks on that slice, which is beneficial for the peak superposition not to exceed the upper limit of the cooling margin. Complementarity This characterizes the complementarity of the two tasks at the amplitude level on that slice. The degree of complementarity is high when one task is in a high-heat phase and the other is in a low-heat phase; the degree of complementarity is low when both are in high-heat phases simultaneously. Specifically, The ratio of the sum of the amplitudes of the two tasks to their respective peak amplitudes can be used to calculate a higher complementarity score for slices with a more balanced amplitude distribution and a more obvious peak-valley misalignment.

[0097] By combining the offset and complementarity of each slice within the candidate time window, a task pair is obtained. Thermal emission complementarity metric within the candidate time window The comprehensive method uses a weighted summation. Each slice's slice-by-slice offset and complementarity are assigned weights. and Calculated as follows: ,

[0098] in The two weighting coefficients can be configured based on the sensitivity of the IDC data center cooling system to peak time misalignment and amplitude balance. When the cooling system responds slowly to instantaneous thermal shocks, the weighting coefficients should be appropriately increased. The proportion of time misalignment characteristics is increased, giving them greater weight in complementary metrics; when cooling capacity margin is tight, the weight is increased. The proportion of tasks is considered, with a focus on selecting task pairs with strong complementary magnitudes.

[0099] The thermal emission complementarity metric for each mission pair As a criterion, a threshold for complementary distribution conditions is set. .when At that time, determine the task to If the complementary distribution condition is met within the candidate time window, it will be included in the set of thermal launch complementary mission pairs. Threshold The threshold setting needs to be considered in conjunction with the actual tightness of the cooling margin and the size of the task queue: if the threshold is too low, a large number of task pairs with similar heat emission distributions will be misjudged as complementary pairs, causing the superimposed heat emission profile to exceed the cooling margin envelope; if the threshold is too high, the number of complementary task pairs will be too small, reducing the scheduling freedom and making it difficult to fully utilize the temporal fluctuation characteristics of the cooling margin. In actual deployment, historical scheduling data can be used to determine the appropriate threshold. Adaptive calibration is performed to keep the measured temperature deviation of the identified complementary task pairs within a safe range after execution.

[0100] In the hot launch complementary mission pair Once determined, the situation where the same task is selected by multiple task pairs needs to be handled. A greedy matching strategy is adopted, based on... Sets from largest to smallest The task pairs are sorted, and the task pairs with the highest complementarity metric are selected and added to the final matching result. Two already matched tasks are removed from the candidate pool to avoid duplicate pairings of the same task. This greedy strategy strikes a balance between computational complexity and matching quality, making it suitable for real-time scheduling scenarios in IDC data centers with large task queues.

[0101] For each thermal emission complementary task pair in the final matching result, they are scheduled to be executed concurrently on adjacent computing nodes within the same cooling partition. This ensures that the heat generated by the two tasks is spatially concentrated within the coverage area of ​​the same cooling unit, facilitating a unified response from the cooling system. Following the time slice order within the candidate time window, the thermal emission intensities of the two tasks to be scheduled in each time slice are added slice by slice to obtain the paired superimposed thermal emission profile of the thermal emission complementary task pair. Let the tasks... In the The thermal emission intensity of each time slice is ,Task The thermal emission intensity of the same slice is The values ​​of the paired superimposed thermal emission profiles on that slice are then... for: ,

[0102] Because the hot launches of the two missions are complementary in terms of time distribution, The fluctuation amplitude on each slice is more stable than the peak value of a single task, and the overall profile tends to be flat, so that the cooling system does not need to be subjected to extreme thermal shock in a short period of time, which helps to reduce cooling energy consumption and extend equipment life.

[0103] The generation of paired superimposed thermal emission profiles provides crucial input for subsequent spatiotemporal alignment and joint optimization. Compared to scheduling the two tasks independently, the peak value of the paired superimposed profile is usually lower than the sum of the peak values ​​of the two independent tasks. This makes it easier to find spatiotemporal placement positions that meet the envelope coverage conditions in the cooling margin time-series slice matrix, thereby improving the feasible domain range and energy efficiency optimization space of the scheduling scheme.

[0104] In one optional implementation, the paired and superimposed thermal emission profiles are taken for the scheduled tasks that have been paired, and the thermal emission timing profiles are taken for the scheduled tasks that have not been paired, and spatiotemporally aligned with the cooling margin timing slice matrix, respectively, including:

[0105] Each task to be scheduled in the task queue is distinguished according to its pairing status. The paired superimposed thermal emission profile is extracted as the placement evaluation profile for the task to be scheduled that has been paired. The thermal emission timing profile is extracted as the placement evaluation profile for the task to be scheduled that has not been paired, thus obtaining a set of placement evaluation profiles for each task to be scheduled.

[0106] Using the cooling partition affiliation corresponding to the placement evaluation contour of each scheduled task as an index, the available cooling margin sequence of the relevant cooling partition dimension corresponding to each scheduled task is extracted from the cooling margin time-series slice matrix to obtain the set of available cooling margin sequences of each scheduled task's cooling partition.

[0107] Using the time slice length of the placement evaluation profile of each task to be scheduled as the alignment window, the placement evaluation profile of each task to be scheduled in the placement evaluation profile set is aligned with the corresponding available cooling margin sequence in the available cooling margin sequence set of the cooling partition at the time slice granularity, so as to obtain the spatiotemporal alignment result of the placement evaluation profile of each task to be scheduled and the corresponding available cooling margin sequence.

[0108] After constructing the complementary and superimposed hot-launch profiles for hot-launch, it is necessary to spatiotemporally align the hot-launch information of each task to be scheduled in the task queue with the cooling margin time-series slice matrix to facilitate subsequent sliding search for feasible spatiotemporal placement locations. The core of spatiotemporal alignment lies in: selecting an appropriate placement evaluation profile based on the pairing status of each task to be scheduled, then extracting the available cooling margin sequence that matches the cooling partition affiliation of the task from the cooling margin time-series slice matrix, and finally achieving precise alignment of the two at the time-slice granularity.

[0109] For each task to be scheduled in the task queue, its state is distinguished based on whether it has completed hot-launch complementary pairing. For tasks that have completed pairing, their hot-launch behavior is presented by the superposition of the two tasks; therefore, the paired superimposed hot-launch profile is extracted as the placement evaluation profile for this task pair. In the paired superimposed hot-launch profile, the hot-launch intensity on each time slice is the slice-by-slice sum of the hot-launch intensities of the two tasks on that slice, fully describing the joint cooling resource requirements of the two tasks when they are executed concurrently on adjacent computing nodes. For tasks that have not completed pairing, their independent hot-launch time sequence profile is directly extracted as the placement evaluation profile. This profile is decomposed according to the internal computing stages and expanded with the predicted hot-launch intensity value of each stage and the corresponding number of time slices, expressing the cooling requirement change process of the task when it is executed independently, slice by slice. After the above distinction processing, a set of placement evaluation profiles for all tasks to be scheduled in the task queue is obtained. Each record in the set contains a task identifier (or task pair identifier), a pairing status flag, and a placement evaluation profile vector expanded by time slices.

[0110] After obtaining the placement evaluation profile set, it is necessary to determine the cooling partition affiliation for each scheduled task, and then extract the corresponding available cooling margin sequence from the cooling margin time-series slice matrix. The cooling partition affiliation is determined based on the following: paired tasks are scheduled to execute concurrently on adjacent compute nodes within the same cooling partition, therefore their cooling partition affiliation is the target cooling partition corresponding to that pair; for scheduled tasks that have not yet been paired, their cooling partition affiliation is determined by the cooling partition of the candidate compute node, and multiple candidate cooling partitions are traversed during the spatiotemporal placement search phase. (Using cooling partition indexes...) Using the cooling headroom time-series slice matrix as the retrieval key, extract the cooling zones. The corresponding available cooling headroom sequence, which is recorded slice by slice along the time axis. The value of covers all time slices from the current moment to the end of the scheduling planning time domain. For completed task pairs, the cooling partition dimension corresponding to their placement evaluation profile is unique, so a single available cooling margin sequence for each cooling partition is extracted. For tasks to be scheduled that have not yet been paired, during the traversal of candidate cooling partitions, the corresponding available cooling margin sequence is extracted for each candidate cooling partition, forming a set of available cooling margin sequences that correspond one-to-one with the set of candidate cooling partitions. After the above extraction operations, a set of available cooling margin sequences for each task to be scheduled is obtained, and each record in the set is associated with a task identifier that is consistent with the corresponding entry in the placement evaluation profile set.

[0111] The key operation of spatiotemporal alignment lies in using the time slice length of the placement evaluation profile as the alignment window to precisely align the placement evaluation profile with the corresponding available cooling margin sequence at the time slice granularity. Let the placement evaluation profile of the task (or task pair) to be scheduled contain a total of... If there are 1 time slice, then the alignment window length is 1. For available cooling margin sequences, their total length is typically much greater than... This covers all time slices within the entire planning time domain. The spatiotemporal alignment process involves sliding a length of [length missing] along the time axis with a step size of 1 on the available cooling margin sequence. The window is used to determine a candidate starting slice position with each slide. The length to be cut off from this position is The available cooled coquanta sequences are matched one-to-one with the placement evaluation contours at the slice granularity, forming a set of spatiotemporal alignment results. In each set of spatiotemporal alignment results, the first... slice With the available cooled co-quantum sequence Each slice corresponds precisely, and the two are compared slice by slice under the same time slice number to determine whether the placement evaluation profile can be covered by the available cooling margin envelope at the candidate starting slice position.

[0112] Regarding the organization of the spatiotemporal alignment results, for each task (or task pair) to be scheduled, the alignment results are indexed by candidate starting slices. For rows, local offset of slices The data is organized into columns to form a two-dimensional alignment matrix. Each element in the matrix contains a pair of records showing the thermal emission intensity value of the evaluation contour placed on the slice and the corresponding available cooling margin value. This organization allows for efficient row-by-row scanning in the subsequent sliding search phase, directly reading the complete alignment pairing data for each candidate starting slice position without repeated extraction and splicing operations, significantly reducing data access overhead during the search process.

[0113] For paired tasks, spatiotemporal alignment is performed on the cooling partition dimension to which the pair belongs. The alignment window length is equal to the total number of time slices of the paired superimposed thermal emission profile. This total number is determined by the number of time slices of the longer task when the two tasks are executed concurrently, and is determined during the paired superimposed thermal emission profile construction phase. For unpaired tasks awaiting scheduling, independent spatiotemporal alignment is performed under each candidate cooling partition. The alignment window length is equal to the total number of time slices of the task's thermal emission time sequence profile. When a task involves multiple cooling partitions (e.g., when a task is deployed across partitions), the available cooling margin sequence needs to be extracted and aligned independently on each relevant cooling partition dimension. Finally, the joint satisfaction of the alignment results of each dimension is used as the basis for judging the feasibility of the candidate position, ensuring that thermal safety constraints are satisfied simultaneously on all relevant cooling partitions.

[0114] After completing the spatiotemporal alignment of the placement evaluation profiles of each task to be scheduled with the corresponding available cooling margin sequence, a structured spatiotemporal alignment result set is obtained. This set provides a directly usable data foundation for subsequent sliding candidate starting slices along the time axis and judging the envelope coverage conditions at each candidate position. This enables the spatiotemporal placement position search process to be carried out efficiently within a unified time slice granularity framework, taking into account the differences between paired and unpaired task pairs in terms of cooling partition affiliation and thermal emission profile morphology, ensuring that the scheduling scheme meets the cooling capacity constraints in both time and space dimensions.

[0115] In one optional implementation, sliding candidate starting slices along the time axis to find a spatiotemporal placement position where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension includes:

[0116] Based on the spatiotemporal alignment results, the starting point of the time axis of the cooling margin time-series slice matrix is ​​taken as the initial candidate starting slice, and the candidate placement interval is defined on the time axis of the cooling margin time-series slice matrix according to the time slice length of the placement evaluation profile of each task to be scheduled.

[0117] The candidate starting slice is moved step by step along the time axis. For each candidate starting slice and the candidate placement interval corresponding to the candidate placement interval, the available cooling margin of each relevant cooling partition dimension in the cooling margin time slice matrix is ​​extracted on the corresponding time slice. The heat emission intensity of the placement evaluation contour on the corresponding time slice is compared on each relevant cooling partition dimension to obtain the time slice and cooling partition dimension envelope coverage determination result corresponding to each candidate starting slice.

[0118] The effective placement condition is that all time-series slice envelope coverage determination results in each relevant cooling partition dimension meet the requirement that the available cooling margin covers the thermal emission intensity of the placement evaluation contour. Candidate starting slices that meet the effective placement condition are extracted from the envelope coverage determination results of all candidate starting slices. The candidate starting slices that meet the effective placement condition are combined with the corresponding cooling partition affiliation information to obtain the spatiotemporal placement position of each task to be scheduled.

[0119] After constructing the cooling margin time-series slice matrix and preparing the placement evaluation profiles for each scheduled task, it is necessary to systematically slide candidate starting slices along the time axis to find effective spatiotemporal placement locations that satisfy thermal safety constraints. The entire sliding search process uses the starting point of the cooling margin time-series slice matrix as the initial candidate starting slice, and calculates the total number of time slices contained in the placement evaluation profile of each scheduled task (or task pair). The length of the candidate placement interval is used to define the corresponding candidate placement interval on the time axis of the cooling margin time slice matrix. For completed paired task pairs, their placement evaluation profile is a paired superimposed thermal emission profile, with a profile length of... The placement evaluation profile is determined by the total number of time slices covered by the two tasks within the candidate time window; for a single task to be scheduled that has not yet been paired, its placement evaluation profile is directly taken from the hot-launch timing profile of that task, and the profile length is... It is obtained by summing the number of time slices in each internal computation stage of the task. The starting point of the candidate placement interval is indexed by the candidate starting slice position. Confirmed, the endpoint of the interval is The requirement is that the entire candidate placement interval does not exceed the effective time axis range of the cooling margin time slice matrix.

[0120] Move the candidate starting slice gradually along the timeline, moving one time slice step at a time, and iterate through all feasible ones in turn. Values. For each given candidate starting slice Within its corresponding candidate placement interval, the available cooling margin for each relevant cooling partition dimension in the time-series slice matrix is ​​extracted on a time-by-time slice basis. Cooling partition index Traverse all relevant cooldown partitions involved in the task (or task pair), time slice index Values , The local offset index for placing the slice within the evaluation contour, with a value range of [value range missing]. to Meanwhile, local offset indices are extracted from the placement evaluation contour. The corresponding thermal emission intensity value, for paired task pairs, is the sum of the paired thermal emission intensities. For an unpaired single task, the thermal emission intensity of that task on the corresponding slice is taken. The two are then compared dimensionally across each relevant cooling partition to determine whether the available cooling margin of the cooling partition on the time slice is not less than the thermal emission intensity of the evaluation contour on the corresponding slice. This yields the time-slice-by-time cooling partition-by-cooling ...

[0121] The core logic of envelope coverage determination lies in: for candidate starting slices It is necessary to process all time slices within the candidate placement interval (i.e. ) and all relevant cooling zone dimensions All meet the available cooling margin. The thermal emission intensity must be no less than that of the evaluation profile placed on the corresponding time slice. If any combination of time slice and cooling zone dimension causes this condition to be unsatisfactory, the candidate starting slice is determined to be invalid and skipped, moving to the next candidate starting slice for further testing. This rigorous, slice-by-slice and dimension-by-dimensional determination mechanism ensures that the final determined spatiotemporal placement location has sufficient cooling capacity in all relevant cooling zones on every time slice throughout the entire mission execution to absorb the corresponding thermal emission load, fundamentally guaranteeing the satisfaction of thermal safety constraints.

[0122] To improve search efficiency, an early termination mechanism can be introduced during the decision-making process: for a given candidate starting slice During the slice-by-slice and dimension-by-dimensional traversal, once a slice offset is detected... With cooling partition The combination does not meet the envelope coverage condition, so the process should be terminated immediately. The subsequent judgment marks the result as not meeting the valid placement conditions and moves on to the next candidate starting slice. This early termination strategy can significantly reduce the number of invalid comparisons and lower the overall search computational cost during periods of tight cooling margin. In addition, for sections in the cooling margin time-series slice matrix that are known to have extremely low cooling margins in certain time slices, they can be pre-marked as prohibited placement intervals. Candidate positions starting from these slices can be skipped directly during the sliding search initialization phase, further shortening the search path.

[0123] After determining the envelope coverage of all candidate starting slices, a set of candidate starting slices that meet the valid placement conditions is extracted from all the determination results. Each candidate starting slice that meets the valid placement conditions... This represents a feasible placement starting point in the time dimension. Combining it with the corresponding cool space allocation information yields a candidate spatiotemporal placement location for the task (or task pair) to be scheduled. The cool space allocation information originates from the association between each task and the cool space determined during the spatiotemporal alignment phase. For paired task pairs, the two tasks are assigned to adjacent computing nodes within the same cool space, therefore their cool space allocation information points to the same cool space. For unpaired single tasks, their cool space allocation information points to the cool space matched for them during the spatiotemporal alignment phase.

[0124] After combining all candidate starting slices that meet the effective placement conditions with the corresponding cooling partition affiliation information, a candidate set of spatiotemporal placement locations for each scheduled task is formed. When a scheduled task has multiple candidate starting slices that meet the effective placement conditions, these candidate starting slices are all retained in the candidate set for further selection of the optimal spatiotemporal placement location in the subsequent joint optimization phase, with the goal of minimizing global energy consumption. When a scheduled task does not have any candidate starting slices that meet the effective placement conditions within the effective time range of the current cooling margin time-series slice matrix, it means that the current cooling capacity configuration cannot safely bear the hot-emission load of the task at any time. The cooling capacity expansion assessment process needs to be triggered, or the task needs to be marked as pending delayed scheduling and the search needs to be re-executed after the cooling margin is updated. Through the above complete sliding search and envelope coverage determination process, a spatiotemporal placement location with a clear time start and cooling partition affiliation is finally generated for each scheduled task (or task pair) in the task queue, providing a feasible domain basis for the generation of subsequent computation-cooling collaborative scheduling schemes.

[0125] In one optional implementation, starting from the aforementioned spatiotemporal placement location and aiming at minimizing global energy consumption and satisfying thermal safety constraints, a collaborative scheduling scheme for IDC data center computing and cooling capacity is generated by jointly optimizing computing resource allocation and cooling capacity response, including:

[0126] Starting from the spatiotemporal placement location, the computational load intensity distribution and cooling demand intensity distribution of each task to be scheduled at the spatiotemporal placement location are extracted. With the minimum global energy consumption as the optimization objective and the thermal safety constraint as the feasible domain boundary, the computational resource allocation dimension and the cooling capacity response dimension are incorporated into the same joint optimization framework to construct a joint optimization objective covering the computational resource allocation dimension and the cooling capacity response dimension.

[0127] Under the constraint of the joint optimization objective, the computational load intensity distribution of each task to be scheduled in the dimension of computational resource allocation is quantified to obtain the computational resource allocation scheme of each task to be scheduled.

[0128] Using the heat load change of each cooling zone corresponding to the computing resource allocation scheme as input, the cooling capacity response of each cooling zone in the cooling capacity response dimension is optimized collaboratively to obtain a cooling capacity response scheme that matches the computing resource allocation scheme.

[0129] Using the convergence criterion that both the computing resource allocation scheme and the cooling capacity response scheme simultaneously satisfy thermal safety constraints and minimize global energy consumption, the computing resource allocation scheme and the cooling capacity response scheme are iteratively verified to obtain the converged computing resource allocation result and cooling capacity response result.

[0130] The converged computing resource allocation results and cooling capacity response results are integrated with the spatiotemporal placement location to generate an IDC data center computing-cooling collaborative scheduling scheme.

[0131] like Figure 2 As shown, the method includes:

[0132] Starting from the spatiotemporal placement location, the computational load intensity distribution and cooling demand intensity distribution of each scheduled task at that location are extracted. The computational load intensity distribution describes the CPU, memory, accelerator, and other computing resource usage required by each scheduled task on its assigned computing node as time slices progress. The cooling demand intensity distribution maps the thermal load magnitude that each cooling partition needs to bear in the corresponding time slice based on the thermal emission time sequence profile or paired superimposed thermal emission profiles. Both types of distribution information are simultaneously incorporated into a joint optimization framework, with the minimum global energy consumption as the optimization objective and meeting thermal safety constraints as the feasible domain boundary, to construct a joint optimization objective function covering both the computational resource allocation dimension and the cooling capacity response dimension.

[0133] The joint optimization objective function consists of two parts: a computational energy consumption term and a cooling energy consumption term. The computational energy consumption term is a weighted sum of the computational resource allocation for each scheduled task across each time slice, with the weighting coefficients reflecting the differences in unit energy consumption among different types of computational resources. The cooling energy consumption term is the sum of the chiller output power for each cooling zone across each time slice, reflecting the total electrical energy consumed by the cooling system to maintain thermal safety. Thermal safety constraints require that the actual heat load of each cooling zone in any time slice does not exceed the upper limit of the heat capacity corresponding to the safe temperature rise threshold of that zone, and that the outlet airflow temperature of each computing node does not exceed the rated thermal design power boundary of the equipment. Integrating the above objective function and constraints forms a joint optimization problem with a clearly defined feasible region.

[0134] Under the constraint of a joint optimization objective, the computational resource allocation dimension is quantitatively solved. For each task to be scheduled, the computational resource requirement corresponding to each stage is inversely calculated based on the predicted thermal emission intensity values ​​of each stage in its thermal emission timeline profile. Specifically, let the task... In the The computational resource allocation vector at the local offset of each slice is: Each component corresponds to the allocation of resources such as the number of CPU cores, memory bandwidth, and accelerator computing power. The range of values ​​is constrained by the upper limit of the physical resources of the computing node, and must also meet the lower bound of the minimum resource requirements for each stage of the task. This is achieved by adjusting the values ​​within the feasible region. An enumeration or gradient-guided search is performed to obtain an initial computing resource allocation scheme that minimizes the energy consumption term on the computing side. For paired task pairs, their computing resource allocation schemes are simultaneously deployed on adjacent computing nodes. The resource allocation vectors of the two nodes are independent of each other in the time slice dimension but share the heat load budget of the same cooling zone.

[0135] Using the heat load changes of each cooling zone corresponding to the computing resource allocation scheme as input, collaborative optimization is performed on the cooling capacity response dimension. The heat load changes of each cooling zone are obtained by transforming the resource allocation vector of each time slice in the computing resource allocation scheme through a heat emission mapping function. The heat emission mapping function maps the computing resource occupancy feature vector to the corresponding heat emission intensity value, and its parameters are derived from the fitting results of historical operating data. Let the cooling zones be... In the The actual heat load on each global time slice is The corresponding cooling capacity response is The goal of optimizing cooling capacity response is to ensure... Under the premise of minimizing the sum of the output power of the refrigeration units in each zone. The coefficient of performance (COP) of the refrigeration units varies with the predicted ambient temperature and the current load rate. During the optimization process, the nonlinear characteristics of COP need to be included in the calculation of the cooling side energy consumption to avoid excessive deviation in energy consumption estimation due to linear approximation. For cooling zones using variable frequency compressors, the cooling capacity response can be continuously adjusted within the rated range; for cooling zones using fixed frequency units, the cooling capacity response is adjusted in units of number, and it needs to be modeled as an integer constraint during optimization.

[0136] After independently solving the initial computational resource allocation scheme and the initial cooling capacity response scheme, the iterative verification phase begins. The convergence criterion for iterative verification is: both the computational resource allocation scheme and the cooling capacity response scheme simultaneously satisfy the thermal safety constraints, and the change in global energy consumption resulting from the joint scheme relative to the previous iteration is less than a preset convergence threshold. In each iteration, using the available cooling margin for each partition determined by the current cooling capacity response scheme as feedback, local adjustments are made to time slices in the computing resource allocation scheme that exceed the cooling margin. Adjustments include reducing the resource allocation for that slice, shifting some computing load to adjacent slices with ample margin, or delaying the task start slice within the task latency tolerance range. The adjusted computing resource allocation scheme re-triggers heat load mapping, obtaining an updated heat load change sequence for each partition, which in turn drives the re-optimization of the cooling capacity response scheme. This process is repeated until both schemes stabilize under the convergence criterion, outputting the converged computing resource allocation result and cooling capacity response result.

[0137] If thermal safety constraints cannot be met during the iteration process, a constraint relaxation mechanism is triggered: this allows for a moderate increase in the upper limit of cooling capacity response within a finite time slice, while simultaneously compressing the computational resource allocation for the corresponding slice to within the safe thermal emission boundary, ensuring that the overall scheduling scheme's feasibility is not compromised by local thermal peaks. The extent of constraint relaxation is constrained by the upper limit of the cooling equipment's rated power and is not allowed to exceed the equipment's physical limits.

[0138] The converged computing resource allocation results, cooling capacity response results, and spatiotemporal placement are integrated to form a complete IDC data center computing-cooling collaborative scheduling scheme. This scheme uses time slices as the granularity, recording the target computing node number, various computing resource allocations, and the number of operating chillers or inverter settings for each scheduled task in each time slice. The collaborative scheduling scheme also includes heat load prediction curves for each cooling zone and global energy consumption estimates for operation and maintenance personnel to review and monitor the scheme's execution. During actual execution, if the measured heat load deviates from the predicted value, local rescheduling can be triggered based on the magnitude of the deviation. This allows for online fine-tuning of the cooling capacity response in the affected time slices without changing the overall spatiotemporal placement, thereby maintaining the continuous effectiveness of computing-cooling collaboration.

[0139] A second aspect of this invention provides an intelligent scheduling system for optimizing energy efficiency in IDC (Internet Data Center) data centers, comprising:

[0140] The cooling margin unit is used to pre-calculate the available cooling margin of each cooling zone based on the periodic change pattern of the operating conditions of the equipment in each cooling zone of the IDC data center and the predicted value of the ambient temperature, and to construct a cooling margin time-series slice matrix.

[0141] The hot emission wheel unit is used to decompose the task to be scheduled according to the internal calculation stage, extract the hot emission intensity and duration of each stage, and obtain the hot emission timing profile of the task to be scheduled.

[0142] The pairing and overlay unit is used to perform pairwise phase analysis on the thermal emission timing profiles of each task to be scheduled in the task queue, identify two pairs of tasks to be scheduled whose thermal emission is complementary in the same candidate time window, and arrange them to be executed concurrently on adjacent computing nodes in the same cooling partition. The thermal emission intensity of the two tasks to be scheduled on each time slice is added slice by slice to obtain the pairing and overlay thermal emission profiles.

[0143] The spatiotemporal alignment unit is used to take the paired superimposed thermal emission profile of the scheduled tasks that have been paired and the thermal emission timing profile of the scheduled tasks that have not been paired, and perform spatiotemporal alignment with the cooling margin timing slice matrix respectively. The candidate starting slice is slid along the time axis to find the spatiotemporal placement position that makes the paired superimposed thermal emission profile or the thermal emission timing profile covered by the available cooling margin envelope in each relevant cooling partition dimension.

[0144] The collaborative scheduling unit is used to generate an IDC data center computing-cooling collaborative scheduling scheme by taking the spatiotemporal placement location as the starting point, aiming at minimizing global energy consumption and meeting thermal safety constraints, and jointly optimizing the allocation of computing resources and cooling capacity response.

[0145] A third aspect of the present invention provides an electronic device, comprising:

[0146] processor;

[0147] Memory used to store processor-executable instructions;

[0148] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0149] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0150] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent scheduling method for optimizing energy efficiency in IDC (Internet Data Center) data centers, characterized in that: include: Based on the periodic change pattern of equipment operating conditions in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin of each cooling zone is pre-calculated on a time-by-time basis, and a time-series slice matrix of cooling margin is constructed. The task to be scheduled is broken down into internal calculation stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled. Pairwise phase analysis is performed on the thermal emission timing profiles of each task to be scheduled in the task queue to identify two pairs of tasks to be scheduled that have complementary thermal emission distributions within the same candidate time window and arrange them to be executed concurrently on adjacent computing nodes in the same cooling zone. The thermal emission intensities of the two tasks to be scheduled on each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles. For tasks that have been paired and scheduled, take the paired superimposed thermal emission profile; for tasks that have not been paired and scheduled, take the thermal emission timing profile. Spatiotemporally align these profiles with the cooling margin timing slice matrix. Slide candidate starting slices along the time axis to find the spatiotemporal placement position where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension. Starting from the aforementioned spatiotemporal placement location, and with the goal of minimizing global energy consumption and meeting thermal safety constraints, a collaborative scheduling scheme for computing and cooling capacity in the IDC data center is generated by jointly optimizing the allocation of computing resources and the response of cooling capacity.

2. The method according to claim 1, characterized in that, Based on the periodic variation patterns of equipment operating conditions in each cooling zone of the IDC data center and the predicted ambient temperature, the available cooling margin for each cooling zone is pre-calculated on a time-series basis, and a time-series slice matrix of cooling margin is constructed, including: The historical operating data of the equipment in each cooling zone of the IDC data center is periodically decomposed, and the periodic reference component and the deviation component of the current time period of each equipment operating parameter are extracted. The periodic reference component is superimposed on the deviation component of the current time period to obtain the periodic change pattern of the equipment operating conditions in each cooling zone. Based on the periodic variation pattern and the measured ambient temperature at the current moment, the predicted ambient temperature value corresponding to each time slice is calculated backward; Based on the positional relationship of time slices, the predicted ambient temperature values ​​of adjacent time slices and non-adjacent time slices are obtained. The upper limit of the available cooling margin of adjacent time slices is corrected by the predicted ambient temperature values ​​of adjacent time slices, and the upper limit of the available cooling margin of non-adjacent time slices is constrained by the predicted ambient temperature values ​​of non-adjacent time slices, thus obtaining the graded corrected upper limit of the available cooling margin for each time slice. Based on the periodic variation pattern and the upper bound of the available cooling margin for graded correction, the available cooling margin of each cooling zone is pre-calculated on a time-by-time slice basis, and the time-series slice matrix of the cooling margin is constructed.

3. The method according to claim 1, characterized in that, The task to be scheduled is broken down into internal computational stages, and the thermal emission intensity and duration of each stage are extracted to obtain the thermal emission timing profile of the task to be scheduled, including: The resource usage characteristics at each moment in the historical execution record of the task to be scheduled are analyzed. The abrupt change position of the resource usage characteristics on the time axis is taken as the boundary of the internal calculation stage. The task to be scheduled is decomposed into multiple sequentially arranged internal calculation stages to obtain the stage boundary sequence of each internal calculation stage. The resource occupancy characteristic intervals of each internal computing stage are defined by the stage boundary sequence, and the resource occupancy characteristics within each internal computing stage interval are aggregated to obtain the resource occupancy benchmark characteristics of each internal computing stage. Using the resource occupancy benchmark features of each internal calculation stage in the historical execution record of the task to be scheduled and the corresponding historical thermal emission record values ​​as samples, a mapping relationship between the resource occupancy benchmark features and thermal emission intensity is established; By taking the resource consumption baseline characteristics of each internal computing stage of the currently scheduled task as input and substituting them into the established mapping relationship, the thermal emission intensity of each internal computing stage can be obtained. The duration of each internal computation stage is calculated based on the start and end boundaries of each internal computation stage in the stage boundary sequence. The thermal emission intensity and duration of each internal computation stage are combined in stage order to obtain the thermal emission timing profile of the task to be scheduled.

4. The method according to claim 1, characterized in that, Pairwise phase analysis is performed on the thermal emission timing profiles of each scheduled task in the task queue to identify two pairs of scheduled tasks with complementary thermal emission distributions within the same candidate time window. These pairs are then arranged to be executed concurrently on adjacent computing nodes within the same cooling zone. The thermal emission intensities of the two scheduled tasks in each time slice are added slice by slice to obtain the paired superimposed thermal emission profiles, including: Extract the extreme positions and extreme amplitudes of the thermal emission intensity from each time slice to obtain the thermal emission phase feature sequence of each task to be scheduled; Using the candidate time window as the alignment benchmark, the thermal emission phase feature sequences of each task to be scheduled in the task queue are compared pairwise. The degree of offset and complementarity of the extreme position of the two tasks to be scheduled are calculated in time slices. The degree of offset and complementarity in time slices are combined into a thermal emission complementarity metric value of the task pair in the candidate time window. Using the thermo-emission complementarity metric of each task pair as a criterion, task pairs whose thermo-emission complementarity metric values ​​satisfy the complementary distribution condition within the same candidate time window are identified, thus obtaining a set of thermo-emission complementary task pairs. Each thermal emission complementary task pair in the set of thermal emission complementary task pairs is arranged to be executed concurrently on adjacent computing nodes within the same cooling partition. According to the time slice order within the candidate time window, the thermal emission intensity of the two tasks to be scheduled in each thermal emission complementary task pair is added slice by slice to obtain the paired superimposed thermal emission profile of each thermal emission complementary task pair.

5. The method according to claim 1, characterized in that, For scheduled tasks that have been paired, the paired superimposed thermal emission profile is taken; for scheduled tasks that have not been paired, the thermal emission timing profile is taken. Spatiotemporal alignment with the cooling margin timing slice matrix is ​​performed, including: Each task to be scheduled in the task queue is distinguished according to its pairing status. The paired superimposed thermal emission profile is extracted as the placement evaluation profile for the task to be scheduled that has been paired. The thermal emission timing profile is extracted as the placement evaluation profile for the task to be scheduled that has not been paired, thus obtaining a set of placement evaluation profiles for each task to be scheduled. Using the cooling partition affiliation corresponding to the placement evaluation contour of each scheduled task as an index, the available cooling margin sequence of the relevant cooling partition dimension corresponding to each scheduled task is extracted from the cooling margin time-series slice matrix to obtain the set of available cooling margin sequences of each scheduled task's cooling partition. Using the time slice length of the placement evaluation profile of each task to be scheduled as the alignment window, the placement evaluation profile of each task to be scheduled in the placement evaluation profile set is aligned with the corresponding available cooling margin sequence in the available cooling margin sequence set of the cooling partition at the time slice granularity, so as to obtain the spatiotemporal alignment result of the placement evaluation profile of each task to be scheduled and the corresponding available cooling margin sequence.

6. The method according to claim 5, characterized in that, Slide candidate starting slices along the time axis to find spatiotemporal placement positions where the paired superimposed thermal emission profile or the thermal emission timing profile is covered by the available cooling margin envelope in each relevant cooling partition dimension, including: Based on the spatiotemporal alignment results, the starting point of the time axis of the cooling margin time-series slice matrix is ​​taken as the initial candidate starting slice, and the candidate placement interval is defined on the time axis of the cooling margin time-series slice matrix according to the time slice length of the placement evaluation profile of each task to be scheduled. The candidate starting slice is moved step by step along the time axis. For each candidate starting slice and the candidate placement interval corresponding to the candidate placement interval, the available cooling margin of each relevant cooling partition dimension in the cooling margin time slice matrix is ​​extracted on the corresponding time slice. The heat emission intensity of the placement evaluation contour on the corresponding time slice is compared on each relevant cooling partition dimension to obtain the time slice and cooling partition dimension envelope coverage determination result corresponding to each candidate starting slice. The effective placement condition is that all time-series slice envelope coverage determination results in each relevant cooling partition dimension meet the requirement that the available cooling margin covers the thermal emission intensity of the placement evaluation contour. Candidate starting slices that meet the effective placement condition are extracted from the envelope coverage determination results of all candidate starting slices. The candidate starting slices that meet the effective placement condition are combined with the corresponding cooling partition affiliation information to obtain the spatiotemporal placement position of each task to be scheduled.

7. The method according to claim 1, characterized in that, Starting from the aforementioned spatiotemporal placement location, and aiming at minimizing global energy consumption and meeting thermal safety constraints, a collaborative scheduling scheme for IDC data center computing and cooling capacity is generated by jointly optimizing computing resource allocation and cooling capacity response, including: Starting from the spatiotemporal placement location, the computational load intensity distribution and cooling demand intensity distribution of each task to be scheduled at the spatiotemporal placement location are extracted. With the minimum global energy consumption as the optimization objective and the thermal safety constraint as the feasible domain boundary, the computational resource allocation dimension and the cooling capacity response dimension are incorporated into the same joint optimization framework to construct a joint optimization objective covering the computational resource allocation dimension and the cooling capacity response dimension. Under the constraint of the joint optimization objective, the computational load intensity distribution of each task to be scheduled in the computational resource allocation dimension is quantified to obtain the computational resource allocation scheme of each task to be scheduled; taking the heat load change of each cooling zone corresponding to the computational resource allocation scheme as input, the cooling capacity response of each cooling zone in the cooling capacity response dimension is collaboratively optimized to obtain a cooling capacity response scheme that matches the computational resource allocation scheme. Using the convergence criterion that both the computing resource allocation scheme and the cooling capacity response scheme simultaneously satisfy thermal safety constraints and minimize global energy consumption, the computing resource allocation scheme and the cooling capacity response scheme are iteratively verified to obtain the converged computing resource allocation result and cooling capacity response result. The converged computing resource allocation results and cooling capacity response results are integrated with the spatiotemporal placement location to generate an IDC data center computing-cooling collaborative scheduling scheme.

8. An intelligent scheduling system for optimizing energy efficiency in IDC data centers, used to implement the method as described in any one of claims 1-7, characterized in that, include: The cooling margin unit is used to pre-calculate the available cooling margin of each cooling zone based on the periodic change pattern of the operating conditions of the equipment in each cooling zone of the IDC data center and the predicted value of the ambient temperature, and to construct a cooling margin time-series slice matrix. The hot emission wheel unit is used to decompose the task to be scheduled according to the internal calculation stage, extract the hot emission intensity and duration of each stage, and obtain the hot emission timing profile of the task to be scheduled. The pairing and overlay unit is used to perform pairwise phase analysis on the thermal emission timing profile of each task to be scheduled in the task queue, identify two pairs of tasks to be scheduled whose thermal emission is complementary in the same candidate time window, and arrange them to be executed concurrently on adjacent computing nodes in the same cooling partition. The thermal emission intensity of the two tasks to be scheduled in each time slice is added slice by slice to obtain the pairing and overlay thermal emission profile. The spatiotemporal alignment unit is used to take the paired superimposed thermal emission profile of the scheduled tasks that have been paired and the thermal emission timing profile of the scheduled tasks that have not been paired, and perform spatiotemporal alignment with the cooling margin timing slice matrix respectively. The candidate starting slice is slid along the time axis to find the spatiotemporal placement position that makes the paired superimposed thermal emission profile or the thermal emission timing profile covered by the available cooling margin envelope in each relevant cooling partition dimension. The collaborative scheduling unit is used to generate an IDC data center computing-cooling collaborative scheduling scheme by taking the spatiotemporal placement location as the starting point, aiming at minimizing global energy consumption and meeting thermal safety constraints, and jointly optimizing the allocation of computing resources and cooling capacity response.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.