Method and system for evaluating load flexibility of data center

Abstract

The embodiment of the invention relates to the technical field of power systems, in particular to a data center load flexibility evaluation method and system, and the method comprises the steps: obtaining an original power load of a data center, and constructing a target function based on the original power load, a power upper limit and a load reduction amount; determining a time slot and a decision variable of a task calculated by the data center, and constructing an upper-layer constraint condition of the target function based on the time slot and the decision variable; obtaining a resource upper limit of the data center, and constructing a lower-layer constraint condition of the objective function based on the resource upper limit and the decision variable; constructing a regression model through the feature extraction function and the regression parameters to obtain a double-layer model; the double-layer model is solved, and the power upper limit and the load reduction amount are confirmed so as to be used for evaluating and describing the load flexibility of the data center. According to the invention, the load flexibility of the data center can be described and evaluated more accurately.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, and particularly relates to a data center load flexibility evaluation method and system. BACKGROUND

[0002] With the rapid development of big data, artificial intelligence (AI) and other technologies, the demand for computing power is rapidly growing worldwide, leading to the continuous rise in the scale and energy consumption of data centers. Although the overall energy consumption is rapidly rising, data centers also exhibit certain power regulation potential. By flexibly scheduling the computing tasks they perform, data centers can dynamically adjust their power load, and thus have the ability to participate in power system demand response.

[0003] If the role of data centers as flexible power resources is to be fully played, the key lies in accurately characterizing their load flexibility, that is, how much load flexibility adjustment capability can be achieved without affecting task completion. By more clearly characterizing load flexibility, the power system can reasonably incorporate it into dispatching and operation to improve the resource scheduling capability of the power system.

[0004] The evaluation and characterization methods in the related art often have the problem of low flexibility level or large fluctuation range, and cannot accurately evaluate under the premise of meeting service guarantee. SUMMARY

[0005] The purpose of the present application is to provide a data center load flexibility evaluation method and system, which can more accurately characterize and evaluate the load flexibility of data centers under the premise of meeting service guarantee.

[0006] To achieve the above purpose, the present application provides the following solutions: In a first aspect, the present application provides a data center load flexibility evaluation method, which comprises: obtaining an original power load of the data center, constructing a target function based on the original power load, a power upper limit and a load reduction amount; confirming time slots and decision variables of computing tasks of the data center, constructing upper constraint conditions of the target function based on the time slots and the decision variables; obtaining a resource upper limit of the data center, constructing lower constraint conditions of the target function based on the resource upper limit and the decision variables; constructing a regression model through a feature extraction function and regression parameters, obtaining a double-layer model according to the regression model after training combined with the target function; solving the double-layer model through the upper constraint conditions and the lower constraint conditions, confirming the power upper limit and the load reduction amount, and using them to evaluate and characterize the load flexibility of the data center.

[0007] In some embodiments, the obtaining the original power load of the data center comprises: confirming a task scheduling algorithm of the data center, and extracting an arrival trajectory of a given computing task in the data center, the arrival trajectory containing an arrival time and task parameter information of the computing task; and confirming the original power load according to the task scheduling algorithm and the arrival trajectory.

[0008] In some embodiments, the time slot is allocated by the data center after each computing task arrives at the data center, the decision variable comprises a time vector of a computing task, the time vector being used to represent a task label and an execution strategy of the computing task, and whether the computing task is executed in the time slot is confirmed by the time vector, the time vector being represented as: wherein, is a variable of 0 or 1, indicating whether the computing task i is executed in the time slot, as the execution strategy of the computing task, and the upper constraint condition comprises a constraint on the execution time of the computing task.

[0009] In some embodiments, the upper constraint condition of the objective function is constructed based on the time slot and the decision variable, comprising: a constraint that a computing task starts to be executed only after a specified time; a constraint that the execution time of the computing task is equal to its duration; a constraint that the execution time of the computing task does not excessively delay based on the time slot setting a deadline and a delay time; and a constraint that a deadline violation probability of a computing task executed in at least one period does not exceed a set value according to the deadline violation probability of the computing task in the period confirmed based on a plurality of arrival trajectories of the computing tasks in the period.

[0010] In some embodiments, the upper constraint condition of the objective function is constructed based on the time slot and the decision variable, comprising: a constraint that a computing task starts to be executed only after a specified time; a constraint that the execution time of the computing task is equal to its duration; a constraint that the execution time of the computing task does not excessively delay based on the time slot setting a deadline and a delay time; and a constraint that a deadline violation probability of a computing task executed in at least one period does not exceed a set value according to the deadline violation probability of the computing task in the period confirmed based on a plurality of arrival trajectories of the computing tasks in the period. wherein, may be any resource-power mapping model, , represents an occupation degree of a computing task to various types of resources; the power consumed by the computing task is calculated based on the resource-power mapping model; and a constraint that the power consumed by all scheduled computing tasks in the same time does not exceed the power upper limit.

[0011] In some embodiments, the constructing the regression model by the feature extraction function and the regression parameter comprises: obtaining a task attribute of a scheduled computing task in the data center, wherein the task attribute at least comprises occupied resource, specified time, and duration of the computing task; confirming a feature vector of the computing task by a feature extraction function; and obtaining the regression model of the computing task scheduling by combining the feature vector and a regression parameter.

[0012] In some embodiments, after the regression model is constructed, the method further comprises: constructing a Lagrange function according to the upper constraint condition and the lower constraint condition: wherein, represents the occupation degree of the computing task to various types of resources, P t cap represents the upper limit of the power, x i,t represents the execution strategy of the computing task; the objective function is constrained based on the Lagrange function to satisfy the KKT condition thereof.

[0013] In some embodiments, the double-layer model is obtained by combining the regression model and the objective function, comprising: embedding the regression model into the objective function to obtain a double-layer model containing regression prediction: wherein, represents the original power load of the data center, is a load reduction, representing the flexible load amount of the data center, P t cap is the upper limit of the power at time t, the double-layer model satisfies the upper constraint, the lower constraint, and the KKT condition; the double-layer model is solved in each time interval T to confirm the upper limit of the power and the load reduction.

[0014] In a second aspect, the present application provides a data center load flexibility evaluation system, comprising: a model construction module, configured to obtain an original power load of the data center, and construct a target function based on the original power load, a power upper limit and a load reduction amount; a constraint module, configured to confirm time slots of calculated tasks of the data center and decision variables, and construct upper constraint conditions of the target function based on the time slots and the decision variables; the constraint module is further configured to obtain resource upper limits of the data center, and construct lower constraint conditions of the target function based on the resource upper limits and the decision variables; the model construction module is further configured to construct a regression model through a feature extraction function and regression parameters; and a data calculation module, configured to solve the target function according to the regression model after training, in combination with the upper constraint conditions and the lower constraint conditions, and confirm the power upper limit and the load reduction amount, so as to evaluate and characterize the load flexibility of the data center.

[0015] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to perform the data center load flexibility evaluation method described above.

[0016] According to the specific embodiments provided in the present application, the following technical effects are disclosed: The present application provides a data center load flexibility evaluation method and system, which firstly constructs a target function based on an original power load, obtains constraint conditions for solving by constructing upper constraint conditions and lower constraint conditions, and then constructs and trains a regression model to combine with the target model to form an optimized double-layer model, and solves in combination with the obtained constraint conditions, so as to obtain real-time power upper limits and load reduction amounts as calculation results. In the calculation process, the influence factors of power capping and task decision can be comprehensively considered, so as to realize more accurate evaluation of load flexibility. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0018] Figure 1 The flowchart of a data center load flexibility evaluation method in an embodiment of the present application.

[0019] Figure 2A comparison chart of the evaluation method and related technology provided by the embodiments of the present application.

[0020] Figure 3 A comparison chart of the calculation efficiency of the evaluation method and related technology provided by the embodiments of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0022] The above purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0023] As shown in the embodiments of the present application, an evaluation method for data center load flexibility is provided, comprising the following steps: Figure 1 S110, obtaining the original power load of the data center, and constructing a target function based on the original power load, the power upper limit and the load reduction amount.

[0024] S120, confirming the time slot and the decision variable of the calculated task of the data center, and constructing the upper constraint condition of the target function based on the time slot and the decision variable.

[0025] S130, obtaining the resource upper limit of the data center, and constructing the lower constraint condition of the target function based on the resource upper limit and the decision variable.

[0026] S140, constructing a regression model through a feature extraction function and a regression parameter, and obtaining a double-layer model according to the trained regression model combined with the target function.

[0027] S150, solving the double-layer model through the upper constraint condition and the lower constraint condition, and confirming the power upper limit and the load reduction amount, so as to evaluate and depict the load flexibility of the data center.

[0028] ​The evaluation method for data center load flexibility in the embodiments of the present application firstly constructs a target function based on the original power load, obtains constraint conditions for solving by constructing upper constraint conditions and lower constraint conditions, and then constructs and trains a regression model to combine with the target model to form an optimized double-layer model, and solves in combination with the obtained constraint conditions to obtain the real-time power upper limit and load reduction amount as the calculation results. In the calculation process, the influence factors of power capping and task decision can be comprehensively considered, so that more accurate evaluation of load flexibility is realized.

[0029] The above steps are described in detail as follows: For step S110, the original power load of the data center is obtained, and a target function based on the original power load, the power upper limit and the load reduction amount is constructed. When the load flexibility needs to be described, an upper model is first designed, and the design goal is to maximize the load flexibility, which is described by the maximum allowed load reduction amount The reduction amount represents the magnitude of the reduction of the power consumption of the data center without violating the service level target.

[0030] In step S110, the original power load of the data center is obtained, including the following processes: Firstly, the task scheduling algorithm of the data center is confirmed, and the arrival trajectory of a given computing task in the data center is extracted, which contains the arrival time and task parameter information of the computing task. Then, the original power load is confirmed according to the task scheduling algorithm and the arrival trajectory. For the arrival trajectory, “arrival” refers to the moment when the data packet is finally received and processed by the target service or processor. “Trajectory” describes the complete process experienced by the data packet in order to “arrive”, which mainly includes path, timing and state. The arrival trajectory of the task and the task scheduling algorithm can confirm the size of the original power.

[0031] The target function constructed is shown as follows: (1) Wherein, represents the original power load of the data center, which is determined by the arrival trajectory of the given computing task in the data center and the original computing task scheduling algorithm. is the maximum allowed load reduction amount of the data center, P t cap is the power upper limit at time t, T represents the time period. Both the power upper limit and the load reduction amount are difficult to obtain directly from the system, so a model needs to be constructed and constrained to solve, so as to confirm the load flexibility of the data center.

[0032] For the upper model, the decision variables also need to be constructed. The time slot of the computing task in the data center and the decision variable are confirmed, and the upper constraint condition of the objective function is constructed based on the time slot and the decision variable. For the above step S120, the time slot is allocated by the data center after each computing task arrives at the data center, and the decision variable includes the time vector of the computing task, which is used to represent the task label and the execution strategy of the computing task, and whether the computing task is executed in the time slot is confirmed through the time vector.

[0033] To solve the problem P t cap , we need to determine the execution time slot of each computing task. In the embodiment of the present application, a computing task i is described by the task attributes, which respectively represent the CPU resource requirement, the memory resource requirement, the I / O resource requirement, the arrival time, the duration and the deadline of the computing task. The execution time of the computing task i is represented by the vector , where is a 0-1 variable, which can be used as an execution decision to indicate whether the computing task i is executed in the time slot.

[0034] For the above step S120, the time slot of the computing task in the data center and the decision variable are confirmed, and the upper constraint condition of the objective function is constructed based on the time slot and the decision variable. Specifically, the following steps are included: Step S121, the execution of the computing task is constrained to start only after the specified time, and the execution decision of the computing task has been described in the above step When the execution strategy indicates that the task is not executed in the time slot, it will be postponed. By postponing non-urgent computing tasks, the data center can reduce its power demand, thereby obtaining a reduced power upper limit curve.

[0035] The upper constraint condition includes: (2) In the above formula, the left side and the right side of the inequality respectively represent the start execution time of the computing task and the arrival time of the computing task. The execution timing of the computing task is constrained by the above formula.

[0036] Step S122, the execution time of the computing task is constrained to be equal to its duration: (3) In the above formula, the left side and the right side of the equality respectively represent the execution time of the computing task and the duration of the computing task calculated by the data center within a time period T. Through the above constraint, the repeated occupation of resources by the computing task can be controlled to be executed within a specified time period.​

[0037] Step S123, based on the time slot setting deadline time and delay time, the execution time of the computing task is not delayed too much. It is represented by the following formula: (4) In the above formula, The deadline time of the task is expected, and γ⋅ΔT represents the delay time. The purpose of the above constraint is to prevent the computing task from being seriously overdue, and the computing task that misses the deadline time is required to complete within γ⋅ΔT time after the deadline time , where γ can be reasonably set according to the expectation of different service level agreement (SLA), and this constraint ensures that the computing task will not be delayed too much.

[0038] Step S124, according to the arrival trajectory of a plurality of computing tasks in at least one period, the deadline violation probability in the period is confirmed, and the deadline violation probability of the computing task execution is not more than a set value.

[0039] For the above process, since the SLA requirement can be described as the deadline violation probability of the computing task execution not more than a given level, the problem can be modeled as the following chance constraint: (5) μ is the deadline violation threshold, and ε is the risk level of the chance constraint optimization, which is set by the SLA requirement, and the chance constraint limits the probability that the proportion of computing tasks violating the deadline exceeds μ to not more than ε. However, the above calculation is only for a single task, and the probability of all computing tasks in a period also needs to be considered: the probability that the computing task violates the deadline comes from the uncertainty of the computing task characteristics, that is, the parameters in the set . Suppose that S computing task arrival trajectories are sampled from the historical data set, and each trajectory s∈{1,2,…,S} represents a possible computing task arrival sequence in the scheduling period. Based on this, the chance constraint of the deadline violation can be represented as: (6) Since the probability of each trajectory is equal, the chance constraint can be rewritten as: (7) Where, represents the deadline violation rate of the scheduling scheme x corresponding to the arrival trajectory s, and the formula ensures that the SLA is satisfied in at least (1−ε)S computing task arrival trajectories. Define whether the computing task i violates the deadline, then we can get: (8) The constraint of the deadline violation rate is given as follows under the given trajectory of the arrival of the computing tasks: (9) In the above formula, the average deadline violation rate of N tasks is given, and a limit is imposed thereon. Thus, the overtime of the computing tasks in the execution process is limited. Through the above steps, the upper constraint condition in the upper model is completed by constraining the task execution strategy.

[0040] On the basis of completing the construction of the upper model, the lower model is constructed. The goal of the lower model is to minimize the number of computing tasks that violate the deadline under the power upper limit Finally, the number of computing tasks that violate the deadline is minimized. In the upper model, for a given trajectory of the arrival of the computing tasks, the power upper limit in each time interval is optimized. The power upper limit in each time interval is converted into a resource constraint (for example, the actual capacity of the active server quantity or available computing resources), and is passed to the lower model as a resource upper limit constraint of the computing task scheduling. Under this constraint, the scheduling of the computing tasks should ensure that the load of the data center does not exceed the limit.

[0041] For the lower model, it is necessary to construct the lower constraint for it according to the upper model. Mainly includes steps S130, obtaining the resource upper limit of the data center, and constructing the lower constraint condition of the objective function based on the resource upper limit and the decision variable.

[0042] Specifically, it includes: S131, obtaining the resource types in the data center and the corresponding resource upper limit.

[0043] S132, the same type of computing resources occupied by all scheduled computing tasks in the same time does not exceed the resource upper limit of the same type of resources. Specifically, it can be embodied by the following formula: (10) Wherein, the resource types can include CPU resources, memory resources, etc. In the above formula, the total CPU resources consumed by all scheduled computing tasks i in time period t must be less than or equal to the total available CPU capacity in this time period, denoted as . Wherein, represents the amount of CPU resources required to execute the computing task i. Similarly, the total memory resources consumed by all scheduled computing tasks i in time period t must be less than or equal to the total available memory capacity in this time period, denoted as . Wherein, represents the amount of memory resources required to execute the computing task i. Similarly, the total I / O resource requirement of all scheduled computing tasks in time period t must not exceed the I / O capacity upper limit.

[0044] The available computing resources of the data center are limited, and the resource usage at any time cannot exceed the upper limit. The resource upper limit of the data center is denoted as The maximum resource amount that the data center can provide is denoted as (11) The above formula indicates that at any time, I The number of computing tasks is denoted as, and the total amount of resources of the computing tasks is represented by summation, which cannot exceed the resource upper limit of the data center.

[0045] In addition, the above steps further include: obtaining the occupation degrees of the computing tasks in the data center to various types of computing resources, constructing a resource-power mapping model in combination with the occupation degrees, calculating the power consumed by the computing tasks based on the resource-power mapping model, and finally constraining the power consumed by all scheduled computing tasks within the same time to be less than the power upper limit.

[0046] The power of the computing task i during execution is denoted as The estimated value of the power can be determined by different computing task power consumption models, which depend on the specific server and the type of computing task: (12) wherein, Any resource-power mapping model can be used. In the above steps, the resource occupation of the computing task is mapped into power consumption. Based on this, the power upper limit constraint can be obtained: (13) wherein, P t cap is the power upper limit at time t t IDC represents the total amount of power consumed by multiple computing tasks.

[0047] A regression model is constructed for obtaining a double-layer model, including the following steps: obtaining the task attributes of the scheduled computing tasks in the data center, wherein the task attributes at least include the occupied resources of the computing tasks, the specified time, and the duration. The feature vector of the computing task is confirmed by a feature extraction function. The regression model of the computing task scheduling is obtained in combination with the feature vector and the regression parameters. The regression model is represented by the following formula: (14) wherein, phi (⋅) represents a feature extraction function that converts the task attributes into a feature vector. The features of the feature vector obtained by the conversion of the task attributes include the required resources of the task, the priority, the data amount, the time difference from the deadline, etc.​ represents the regression parameter, which is trained separately in each time period. The variable before the regression model equation represents the execution strategy of the computing task in the data center.

[0048] After the regression model is constructed, it needs to be trained. The generation process of the training data is as follows: first, different load reduction amounts are sampled to construct multiple training scenarios, then the optimal scheduling is solved in each scenario, and finally the task features and optimal scheduling results are combined to form a sample pair to form a data set for training the regression model.

[0049] When training the regression model for each time period t, the regularized mean square error is optimized: (15) where the right side of the equation is the sum of two terms. The first term represents the mean square error between the model prediction result and the optimal ILP (integer linear programming) solution, which is used to measure the prediction accuracy. The second term is a regularization term, which is used to suppress the model parameters from being too large and prevent overfitting. λ is the regularization coefficient, which is used to balance the trade-off between prediction error and model complexity. Through the above training method, the prediction result gradually tends to decay, and the regression model of the training sum is obtained.

[0050] In some embodiments, the constraint condition also includes the KKT condition: According to the upper constraint condition and the lower constraint condition, a Lagrange function is constructed for constructing an integrated framework, and the Lagrange function is as follows: (16) The constraint target function satisfies the KKT condition of the Lagrange function, and the KKT condition includes: (17) wherein, represents the occupation degree of the computing task to various types of resources, P t cap represents the upper limit of the power, x i,t represents the execution strategy of the computing task.

[0051] Finally, the regression model is embedded into the target function to obtain a double-layer model containing regression prediction: The constraints include: 1, , i.e. the upper constraint condition in the foregoing based on the task execution decision.

[0052] 2, , , i.e. the lower constraint condition in the foregoing, constrained by the upper limit of power and the upper limit of resources.

[0053] 3, the constraint of KKT condition.

[0054] Wherein, represents the original power load of the data center, is the load reduction, representing the flexible load amount of the data center, P t cap is the power upper limit at time t , the bi-level model satisfies the upper constraint, the lower constraint and the KKT condition, and the bi-level model is solved in each time interval T to confirm the power upper limit and the load reduction.

[0055] In the above manner, the maximum load reduction capability of the data center at each period, i.e. the flexibility of the load center, can be described.

[0056] The evaluation method of the load flexibility of the data center in the embodiment of the application first constructs a target function based on the original power load, obtains the constraint condition for solving by constructing the upper constraint condition and the lower constraint condition, and then constructs and trains the regression model to combine with the target model to form an optimized bi-level model. The constraint condition obtained is combined to solve, so as to obtain the real-time power upper limit and the load reduction as the calculation result. In the calculation process, the influence factors of the power cap and the task decision can be considered comprehensively, so as to realize more accurate evaluation of the load flexibility.

[0057] Please refer to Figure 2 , which shows the time-varying power flexibility realized by three different methods in a day. Specifically, the data-driven flexibility assessment (Data-driven Flexibility Assessment, DFA for short) method, the virtual capacity curve (Virtual Capacity Curve, VCC for short) method, and the flexibility description method in the embodiment of the application.

[0058] Figure 2 In the figure, the solid line represents the method in the embodiment of the application, the dotted line represents the DFA method, and the dashed line represents the VCC method. The horizontal coordinate in the figure represents time, and the vertical coordinate represents the flexibility of the data center. Figure 2The data center flexibility can be characterized with different methods. It can be seen that the method proposed in the embodiments of the present application is superior to the two benchmark methods in flexibility implementation: the DFA method generates a relatively smooth flexibility curve due to its inherent assumption of periodic workload and limited time adjustability, but the overall flexibility level is significantly lower. The VCC method exhibits higher peak flexibility in some periods (especially between 11:00 and 15:00), which is due to its carbon intensity driven power reduction strategy. However, this method does not coordinate the scheduling constraints at the computing task level, resulting in aggressive capacity reduction during high load periods, which in turn leads to system instability and higher risk of task deadline violation.

[0059] In contrast, the method proposed in the embodiments of the present application, based on the double-layer model, explicitly coordinates the power cap decision and the feasibility of real-time computing task scheduling under strict service level agreement (SLA) constraints, thereby ensuring maximum power flexibility without violating service guarantees, with better accuracy to support the data center's participation in demand response of the power system.

[0060] Please refer to Figure 3 for the comparison of the evaluation method of flexibility in the embodiments of the present application and the model solving in the prior art. Among them, the horizontal coordinate represents the number of computing tasks in the data center, and the vertical coordinate represents the calculation time of solving the model. Figure 3 The dots and blocks in correspond to the related art and the method of the present application, respectively. Figure 3 The results in show that as the arrival rate of computing tasks increases, the calculation time of directly solving the double-layer model increases exponentially and is significantly longer than the regression-based method - which may eventually take several hours. For example, when the arrival rate of computing tasks reaches 20,000 per hour, the solving time exceeds 2.8 hours, which makes it unsuitable for real-time applications.

[0061] In contrast, the regression-based method proposed in the embodiments of the present application always maintains the solving time within tens of seconds even in the case of 25,000 computing tasks per hour. The results show that this method can significantly improve the computing efficiency and is suitable for real-time demand flexibility quantification of data centers. Its advantage lies in that the regression model can achieve fast reasoning through simple feature mapping, thereby avoiding the time-consuming process of solving large-scale double-layer problems in real-time scenarios.

[0062] It can be determined through the above comparison that the data center load flexibility evaluation method provided in the embodiments of the present application can comprehensively consider the influence factors of power capping and task decision in the calculation process, thereby realizing more accurate evaluation of load flexibility. Moreover, the method of the embodiments of the present application adopts a regression-based method to calculate the double-layer model in the process of solving and calculating, which can significantly improve the efficiency of evaluation and characterization.

[0063] The embodiments of the present application also provide a data center load flexibility evaluation system. The data center load flexibility evaluation system comprises a model construction module, a constraint module, and a data calculation module. The model construction module is configured to obtain an original power load of a data center, and construct a target function based on the original power load, a power upper limit, and a load reduction amount. The constraint module is configured to confirm a time slot and a decision variable of a calculated task of the data center, and construct an upper constraint condition of the target function based on the time slot and the decision variable. The constraint module is also configured to obtain a resource upper limit of the data center, and construct a lower constraint condition of the target function based on the resource upper limit and the decision variable. The model construction module is also configured to construct a regression model through a feature extraction function and a regression parameter. The data calculation module is configured to solve the target function according to the regression model after training, in combination with the upper constraint condition and the lower constraint condition, to confirm the power upper limit and the load reduction amount, so as to evaluate and characterize the load flexibility of the data center. The embodiments of the present application also provide a computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to perform the data center load flexibility evaluation method described above.

[0064] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use, and processing of related data need to comply with relevant regulations.

[0065] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to a memory, a database or other medium used in the embodiments provided in the present application can include at least one of a non-volatile and a volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive random access memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric random access memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc.

[0066] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0067] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0068] The principles and implementation modes of the present application are described by applying specific examples herein, and the above-mentioned embodiments are only used to help understand the method and its core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range can be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A method for evaluating data center load flexibility, characterized in that, The methods for assessing data center load flexibility include: Obtain the original power load of the data center and construct an objective function based on the original power load, the power limit, and the load reduction amount; Identify the time slots and decision variables for the tasks calculated by the data center, and construct the upper-level constraints of the objective function based on the time slots and decision variables; Obtain the resource limit of the data center, and construct the lower-level constraints of the objective function based on the resource limit and the decision variables; A regression model is constructed by using feature extraction functions and regression parameters, and a two-layer model is obtained by combining the trained regression model with the objective function. The two-layer model is solved by the upper-layer constraints and the lower-layer constraints to confirm the power limit and the load reduction.

2. The data center load flexibility assessment method according to claim 1, characterized in that, The process of obtaining the raw power load of the data center includes: The task scheduling algorithm of the data center is confirmed, and the arrival trajectory of a given computing task in the data center is extracted. The arrival trajectory includes the arrival time and task parameter information of the computing task. The original power load is confirmed based on the task scheduling algorithm and the arrival trajectory.

3. The data center load flexibility assessment method according to claim 1, characterized in that, The time slot is allocated by the data center after each computing task arrives at the data center. The decision variable includes the time vector of the computing task, which represents the task label and execution strategy of the computing task. The time vector is used to determine whether the computing task is executed within the time slot. The time vector is represented as follows: in, It is a variable of 0 or 1, indicating whether the computation task i is executed within the time slot, as the execution strategy of the computation task. The upper-level constraints include constraints on the execution time of the computation task.

4. The data center load flexibility assessment method according to claim 3, characterized in that, The upper-level constraints for constructing the objective function based on the time slot and the decision variables include: The constraint calculation task will only begin execution after the specified time has elapsed; The execution time of the computation task is constrained to be equal to its duration; The deadline and delay time are set based on the time slot to constrain the execution time of the computing task from being excessively delayed. The probability of deadline default within a period is determined based on the arrival trajectories of multiple computation tasks within at least one period, and the probability of deadline default for the execution of computation tasks is constrained to not exceed a set value.

5. The data center load flexibility assessment method according to claim 1, characterized in that, The step of obtaining the resource limit of the data center and constructing the lower-level constraints of the objective function based on the resource limit and the decision variables includes: Obtain the types of resources and their corresponding upper limits within the data center; Within the same time period, the amount of computing resources of the same type used by all scheduled computing tasks shall not exceed the resource limit of that type. Obtain the resource occupancy of various types of computing resources by computing tasks in the data center, and construct a resource-power mapping model based on the resource occupancy. in, It can be any resource-power mapping model. , , R i I / O This indicates the resource consumption of the corresponding type by the computing task; The power consumed by the computational task is calculated based on the resource-power mapping model. The power consumed by all scheduled computing tasks within the same time period shall not exceed the power limit.

6. The data center load flexibility assessment method according to claim 5, characterized in that, The construction of the regression model through feature extraction functions and regression parameters includes: Obtain the task attributes of the scheduled computing tasks in the data center, wherein the task attributes include at least the resources used by the computing task, the specified time, and the duration; The feature vector of the computational task is identified through a feature extraction function; The regression model for scheduling the computational task is obtained by combining the feature vector and the regression parameters.

7. The data center load flexibility assessment method according to claim 5, characterized in that, After constructing the regression model, the method further includes: Construct the Lagrangian function based on the upper-level constraints and the lower-level constraints: in, P represents the resource consumption of various types by the computational task. t cap x represents the upper limit of power. i,t The execution strategy for the computation task is defined; the objective function is constrained to satisfy its KKT conditions.

8. The method for evaluating data center load flexibility according to claim 7, characterized in that, The process of obtaining a two-layer model by combining the trained regression model with the objective function includes: Embedding the regression model into the objective function yields a two-layer model that includes regression predictions: in, This represents the raw power load of the data center. The load reduction amount represents the flexible load capacity of the data center. P t cap It is a moment t The power upper limit, the two-sided model satisfies the upper constraint, the lower constraint and the KKT condition; The two-layer model is solved at each time interval T to determine the power limit and the load reduction.

9. A data center load flexibility assessment system, characterized in that, The data center load flexibility assessment system includes: The model building module is used to obtain the original power load of the data center and construct an objective function based on the original power load, the power limit, and the load reduction amount. The constraint module identifies the time slots and decision variables of the tasks calculated by the data center, and constructs upper-level constraints for the objective function based on the time slots and decision variables; the constraint module is also used to obtain the resource limit of the data center, and construct lower-level constraints for the objective function based on the resource limit and decision variables. The model building module is also used to build a regression model using feature extraction functions and regression parameters; The data calculation module is used to solve the objective function based on the trained regression model and the upper-level constraints and the lower-level constraints, and to determine the power limit and the load reduction amount, so as to evaluate and characterize the load flexibility of the data center.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the data center load flexibility assessment method according to any one of claims 1-8.

Images