Green computing power scheduling system and method based on multi-objective optimization

Abstract

The application provides a green computing power scheduling system and method based on multi-objective optimization, relates to the technical field of green computing power scheduling, and comprises the following steps: receiving a plurality of computing tasks, obtaining task attribute information corresponding to each computing task, and the task attribute information comprising task type information; collecting GPU power and temperature data of each data center in a data center cluster, calculating the use efficiency value of each data center power supply, and obtaining the electricity price information and carbon intensity data of each period in the area covered by each data center from an application programming interface connected to the power grid; based on the task type information and the temperature data, dynamically adjusting the GPU power of each data center to obtain an adjustment result, scheduling and optimizing each computing task to obtain an optimization result, and then constructing a decision vector, and finally performing computing power scheduling through a multi-objective optimization function and an adaptive scheduling strategy, so that the application can realize green and efficient computing power scheduling.

Description

Technical Field

[0001] This application relates to the technical field of green computing power scheduling, and in particular to a green computing power scheduling system and method based on multi-objective optimization. Background Technology

[0002] In computing-intensive scenarios such as data center clusters and AI training, green computing scheduling is the key to achieving efficient and low-consumption operation. As energy consumption and carbon emission constraints become increasingly stringent, the application demand for green computing scheduling methods based on multi-objective optimization continues to grow and has broad prospects.

[0003] Currently, most common multi-objective computing power scheduling technologies focus on optimizing task completion time and total energy consumption. Some solutions combine electricity price differences for time-of-day scheduling and attempt to reduce operating costs. A few technologies involve power adjustment of individual devices.

[0004] However, these existing technologies often overlook the impact of carbon intensity fluctuations and differences in energy efficiency among different data centers, making it difficult to achieve synergistic optimization of energy consumption, carbon emissions, and costs while ensuring the normal completion of tasks, resulting in limited overall effectiveness of green scheduling. Summary of the Invention

[0005] The purpose of this application is to provide a green computing power scheduling system and method based on multi-objective optimization, so as to solve the problem that computing power scheduling in the prior art is difficult to take into account both multi-dimensional green objectives and task execution requirements.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a green computing power scheduling method based on multi-objective optimization, comprising:

[0007] Receive multiple computing tasks and obtain task attribute information corresponding to each computing task, wherein the task attribute information includes task type information;

[0008] Collect GPU power and temperature data from each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data for each time period in the area covered by each data center from the application programming interface connected to the power grid.

[0009] Based on the task type information and the temperature data, the GPU power of each data center is dynamically adjusted to obtain the adjustment result. Based on the carbon intensity data, the utilization efficiency value, and the electricity price information, the scheduling of each computing task is optimized to obtain the optimization result.

[0010] Based on the adjustment and optimization results, a decision vector is constructed, which is used to represent the scheduling scheme to be determined.

[0011] By combining a preset multi-objective optimization function and adopting an adaptive scheduling strategy, the decision vector is subjected to computing power scheduling to obtain the scheduling result. The multi-objectives include: task completion time, total energy consumption, carbon emissions, and electricity cost.

[0012] Optionally, the step of dynamically adjusting the GPU power of each data center based on the task type information and the temperature data to obtain the adjustment result, and optimizing the scheduling of each computing task based on the carbon intensity data, the utilization efficiency value, and the electricity price information to obtain the optimization result, includes:

[0013] Based on the preference tags corresponding to the task type information, configure the corresponding energy efficiency strategy for the computation task to obtain the strategy allocation result. The task type information is an emergency task, a batch processing task, or a training task.

[0014] Based on the strategy allocation results and the pre-established performance and power consumption curve model, combined with the temperature data, the GPU operating voltage and frequency are coordinated and adjusted to obtain an adjustment result that meets temperature safety constraints.

[0015] Based on the carbon intensity data and the electricity price information, the calculation tasks are scheduled in time to obtain the scheduling results. Based on the utilization efficiency value, the calculation tasks are assigned nodes to obtain the node assignment results.

[0016] The time scheduling results and the node allocation results are integrated to form an optimized result.

[0017] Optionally, the step of coordinating the adjustment of GPU operating voltage and frequency based on the strategy allocation result and a pre-established performance-power curve model, combined with the temperature data, to obtain an adjustment result that meets temperature safety constraints includes:

[0018] Based on the strategy allocation results and the pre-established performance and power consumption curve model, determine the combination of adjustment parameters for GPU operating voltage and frequency;

[0019] Based on the temperature data and preset temperature safety constraints, the combination of adjustment parameters is constrained, and the operating voltage and frequency of the GPU are synchronously adjusted using the constrained combination of adjustment parameters to obtain the adjustment result.

[0020] Optionally, the step of scheduling the computational tasks based on the carbon intensity data and the electricity price information to obtain a scheduling result, and then performing node allocation processing on the computational tasks based on the utilization efficiency value to obtain a node allocation result, includes:

[0021] Based on the periodic variation patterns of the carbon intensity data and the electricity price information, the distribution characteristics of the target time period are identified to obtain the time period distribution information;

[0022] Based on the time period distribution information, the execution time of the computing tasks is arranged to obtain the time arrangement result;

[0023] Select target data centers from the data center cluster that have a power efficiency value lower than a preset power efficiency threshold and a cooling efficiency higher than a preset cooling efficiency threshold.

[0024] Based on the load status of the target data center, a corresponding target data center is assigned to each computing task to obtain the node allocation result.

[0025] Optionally, the step of combining a preset multi-objective optimization function and employing an adaptive scheduling strategy to perform computational power scheduling on the decision vector to obtain a scheduling result includes:

[0026] Based on the latest completion time in the task attribute information, determine the completion time constraint; based on the service quality requirements, determine the service quality constraint; based on the resource requirements in the task attribute information, determine the resource constraint; and based on the temperature data, determine the temperature safety constraint.

[0027] The completion time constraint, the service quality constraint, the resource constraint, and the temperature safety constraint are integrated to obtain a set of constraint conditions;

[0028] Based on task type information and preference labels, the weight coefficients of each objective in the preset multi-objective optimization function are configured differently to obtain an adaptive scheduling strategy.

[0029] Based on the set of constraints and the adaptive scheduling strategy, the decision vector is optimized to obtain multiple candidate decision vectors;

[0030] The Pareto optimization method is used to select the target decision vector from all the candidate decision vectors, and the target decision vector is converted into the corresponding scheduling result.

[0031] Optionally, the step of differentiating the weight coefficients of each objective in a preset multi-objective optimization function based on task type information and preference labels to obtain an adaptive scheduling strategy includes:

[0032] Based on task type information, determine the number of tasks for each task type;

[0033] Based on the number of tasks and corresponding preference labels for each task type, a global optimization strategy is used to differentiate the weight coefficients of each objective in the multi-objective optimization function to obtain a weight allocation scheme.

[0034] The weight allocation scheme is associated with the functions of each objective in the multi-objective optimization function to form an adaptive scheduling strategy.

[0035] Optionally, constructing a decision vector based on the adjustment and optimization results includes:

[0036] Extract the adjustment parameters corresponding to the GPU operating voltage and frequency from the adjustment results, and extract the node allocation parameters and time scheduling parameters of the computing tasks from the optimization results;

[0037] According to the preset encoding rules, the adjustment parameters corresponding to the GPU operating frequency and voltage, as well as the node allocation parameters and time scheduling parameters of the computing tasks, are combined to construct a decision vector.

[0038] Secondly, this application provides a green computing power scheduling system based on multi-objective optimization, comprising:

[0039] The acquisition module is used to receive multiple computing tasks and acquire task attribute information corresponding to each computing task, wherein the task attribute information includes task type information.

[0040] The data acquisition module is used to collect GPU power and temperature data from each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data for each time period in the area covered by each data center from the application programming interface connected to the power grid.

[0041] The adjustment module is used to dynamically adjust the GPU power of each data center based on the task type information and the temperature data to obtain the adjustment result; and to optimize the scheduling of each computing task based on the carbon intensity data, the utilization efficiency value, and the electricity price information to obtain the optimization result.

[0042] A construction module is used to construct a decision vector based on the adjustment and optimization results, wherein the decision vector is used to represent the scheduling scheme to be determined;

[0043] The scheduling module is used to combine a preset multi-objective optimization function and adopt an adaptive scheduling strategy to perform computing power scheduling on the decision vector to obtain the scheduling result. The multi-objectives include: task completion time, total energy consumption, carbon emissions, and electricity cost.

[0044] Thirdly, this application provides an electronic device, comprising:

[0045] Memory, used to store computer programs;

[0046] A processor, used to execute the computer program to implement the steps of the green computing power scheduling method based on multi-objective optimization as described in the first aspect above.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the green computing power scheduling method based on multi-objective optimization as described in the first aspect above.

[0048] The green computing power scheduling method based on multi-objective optimization provided in this application has the following beneficial effects:

[0049] This application first receives multiple computing tasks and obtains attribute information including task type information to clarify the core characteristics of the tasks, which can provide a basis for subsequent targeted scheduling; then it collects GPU power and temperature data and calculates power efficiency values, and then obtains electricity price and carbon intensity data to comprehensively understand the relevant information of computing power operation, which can provide data support for scheduling decisions.

[0050] Then, the GPU power is dynamically adjusted based on task type and temperature data, and the computational task is optimized by combining carbon intensity, utilization efficiency value and electricity price scheduling to adapt to task requirements and running status, thereby reducing ineffective energy consumption. After that, a decision vector is constructed based on the adjustment and optimization results to visualize the scattered scheduling basis and form a clear and optimizable undetermined scheduling scheme. Finally, the decision vector is scheduled by combining a multi-objective optimization function and an adaptive scheduling strategy to balance task completion time, energy consumption, carbon emissions and cost, so as to obtain the comprehensive optimal scheduling result.

[0051] Furthermore, this application enables GPU adjustment to better meet task requirements by configuring energy efficiency strategies based on different types. Combined with temperature data, voltage and frequency can be adjusted in a coordinated manner to ensure safe operation of the device. At the same time, through precise matching of time and nodes, task execution can be more aligned with green goals, thereby improving the targeting and greening level of scheduling and further optimizing energy consumption, carbon emissions, and cost control. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating a green computing power scheduling method based on multi-objective optimization provided in this application embodiment;

[0054] Figure 2 A schematic diagram illustrating a specific implementation of a green computing power scheduling method based on multi-objective optimization provided in this application embodiment;

[0055] Figure 3 A schematic diagram of the structure of a green computing power scheduling system based on multi-objective optimization provided in an embodiment of this application;

[0056] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0057] In the current environment where computing power-intensive scenarios are increasing, the need to control energy consumption and carbon emissions is becoming more and more urgent. While existing multi-objective computing power scheduling technologies can achieve optimization in some dimensions, they mostly focus on task completion time and single energy consumption indicators, while ignoring the dynamic fluctuations of grid carbon intensity and the differences in energy efficiency among different data centers. This makes it difficult to simultaneously optimize energy consumption, carbon emissions, and operating costs while ensuring the normal execution of tasks, and the overall effect of green scheduling fails to meet the needs of practical applications.

[0058] To address this issue, this application proposes a green computing power scheduling method based on multi-objective optimization. This method first acquires information about the type of computing task, then comprehensively collects data on data center equipment operation, energy efficiency, and electricity prices and carbon intensity from the power grid. Subsequently, it adjusts equipment power and optimizes task execution time and deployment nodes accordingly. Finally, it completes scheduling through a multi-objective optimization strategy. This solution integrates multi-dimensional key data and adapts to different task requirements for precise scheduling, ensuring successful task completion while effectively balancing energy consumption, carbon emissions, and costs. This successfully solves the core problem in existing technologies where computing power scheduling struggles to simultaneously address multi-dimensional green objectives and task execution needs.

[0059] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] The core of this application is to provide a green computing power scheduling method based on multi-objective optimization, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0061] S101. Receive multiple computing tasks and obtain task attribute information corresponding to each computing task, wherein the task attribute information includes task type information.

[0062] Among them, task attribute information refers to the set of information describing the core characteristics of the computing task, which includes key content such as task type information, and can provide a basis for subsequent targeted scheduling strategies; task type information refers to the classification identifier of computing tasks, used to distinguish tasks with different execution requirements, such as urgent tasks that need to be prioritized, batch processing tasks that can be flexibly scheduled, and training tasks that need to be executed in stages.

[0063] In step S101, multiple computing tasks submitted from different business scenarios or users are first received. Then, key information corresponding to each computing task is collected through common channels such as task submission interface and system registration information. Subsequently, the core content related to scheduling is selected and integrated into task attribute information, while the task type information is clarified, and finally, structured task basic data is formed.

[0064] S102. Collect GPU power and temperature data of each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data of each time period in the area covered by each data center from the application programming interface connected to the power grid.

[0065] Among them, the power utilization efficiency value is an indicator that measures the energy utilization efficiency of a data center. It reflects how much of the total electrical energy consumed by the data center is actually used for the operation of computing equipment, and can help to intuitively judge the energy efficiency of different data centers.

[0066] Carbon intensity data refers to the amount of carbon emissions per unit of electricity. It is used to reflect the environmental friendliness of electricity in different time periods and regions, and thus provides a core reference for low-carbon-oriented dispatching decisions.

[0067] In step S102, the power data of the GPUs in each data center, as well as the temperature data of the surrounding environment and the chip itself, are captured in real time by the equipment monitoring components deployed in the data center. Then, the total energy consumption record of the data center and the actual energy consumption data of the computing equipment are combined to obtain the power utilization efficiency value of each data center through preset calculation logic. At the same time, the electricity price standard of each data center area at different times and the corresponding carbon intensity data are obtained in batches through a dedicated data interface connected to the power grid system. Finally, all the collected and calculated data are classified and organized to form a standardized scheduling data set.

[0068] It should be noted that the specific content of the preset calculation logic is not limited in the embodiments of this application, and can be set accordingly according to the actual situation.

[0069] This application integrates data related to equipment operating status, data center energy efficiency, external electricity costs, and environmental attributes, providing multi-dimensional data support for subsequent precise adjustment of equipment operating status and optimization of task execution arrangements, thereby ensuring the scientific nature and pertinence of scheduling decisions.

[0070] S103. Based on the task type information and the temperature data, dynamically adjust the GPU power of each data center to obtain the adjustment result. Based on the carbon intensity data, the utilization efficiency value, and the electricity price information, optimize the scheduling of each computing task to obtain the optimization result.

[0071] The GPU power is calculated from the GPU frequency and voltage.

[0072] In one specific implementation, such as Figure 2 As shown, step S103 includes:

[0073] Step 1031: Configure the corresponding energy efficiency strategy for the calculation task according to the preference tag corresponding to the task type information, and obtain the strategy allocation result. The task type information is an emergency task, a batch processing task, or a training task.

[0074] Among them, preference tags refer to the requirement identifiers that match the type of computing task, which are used to clarify the core orientation of the task in terms of performance, energy efficiency or execution method; energy efficiency strategies refer to the equipment operation and scheduling rules formulated for different task types, which are used to establish an adaptation relationship between task requirements and green goals; strategy allocation results refer to the set of correspondences between task type, preference tags and energy efficiency strategies, which can provide a clear basis for subsequent GPU adjustment.

[0075] In step 1031, firstly, according to the three types of computing tasks—urgent tasks, batch processing tasks, and training tasks—corresponding preference tags are matched for each type. Then, appropriate energy efficiency strategies are configured for different preference tags. Urgent tasks correspond to a performance-first strategy that prioritizes execution speed, batch processing tasks correspond to an energy-first strategy that prioritizes reducing consumption, and training tasks correspond to a segmented execution strategy that allows for phased execution. Finally, a strategy allocation result containing the correspondence between task type, preference tag, and energy efficiency strategy is formed.

[0076] For example, computing tasks are categorized into three types: urgent tasks, batch processing tasks, and training tasks. Performance-priority, energy-efficiency-priority, and segmented execution preference labels are matched to the three types of tasks respectively. This results in a strategy allocation result: performance-priority strategy for urgent tasks, energy-efficiency-priority strategy for batch processing tasks, and segmented execution strategy for training tasks.

[0077] Step 1032: Based on the strategy allocation result and the pre-established performance and power consumption curve model, and combined with the temperature data, the GPU operating voltage and frequency are adjusted in a coordinated manner to obtain an adjustment result that meets the temperature safety constraints.

[0078] Among them, the performance and power consumption curve model refers to the model that describes the relationship between GPU performance indicators and operating voltage, frequency and power consumption, and is used to quickly query the performance and power consumption levels corresponding to different parameter combinations.

[0079] Step 1032 may specifically include the following steps:

[0080] Step a1: Based on the strategy allocation results and the pre-established performance and power consumption curve model, determine the combination of adjustment parameters for GPU operating voltage and frequency.

[0081] In step a1, the computing speed and actual power consumption data of the GPU under different voltage and frequency combinations are collected first. Then, the correlation between the data is established through data fitting to obtain a curve model of performance and power consumption. Subsequently, based on the obtained strategy allocation results, the candidate range of voltage and frequency under the corresponding strategy is queried from the model. Then, the adjustment parameter combination that meets the requirements is selected according to the minimum performance requirements of the task. The performance priority strategy corresponds to the standard voltage and frequency parameters, and the energy efficiency priority strategy corresponds to the reduced voltage and frequency parameters.

[0082] For example, by using the performance-power curve model to query the voltage and frequency candidate ranges for the corresponding strategies of the three types of tasks, and by combining the minimum performance requirements of the tasks to select the adjustment parameter combinations, the adjustment parameter combinations corresponding to the emergency task are obtained as follows: voltage 1.0V and frequency 1.5GHz, batch processing task is voltage 0.9V and frequency 1.2GHz, and training task is voltage 0.98V and frequency 1.3GHz.

[0083] Step a2: Based on the temperature data and the preset temperature safety constraints, the combination of adjustment parameters is constrained, and the operating voltage and frequency of the GPU are synchronously adjusted using the constrained combination of adjustment parameters to obtain the adjustment result.

[0084] Temperature safety constraints refer to the maximum temperature standards set to prevent GPU chips from overheating and being damaged or experiencing performance degradation; the adjustment parameter combination refers to the specific combination of GPU operating voltage and frequency, which directly determines the GPU's performance and power consumption.

[0085] In step a2, the chip junction temperature value is first obtained from the temperature data. Then, the obtained combination of adjustment parameters is substituted into the preset correlation formula between parameters and junction temperature to calculate the estimated junction temperature value. If the estimated junction temperature exceeds the preset temperature safety constraint, the frequency or voltage is gradually reduced to regenerate the combination of adjustment parameters and calculate the estimated junction temperature again until the estimated junction temperature meets the temperature safety constraint. Finally, the finally determined combination of adjustment parameters is used to synchronously adjust the GPU's operating voltage and frequency to obtain the adjustment result.

[0086] For example, the formula for calculating the estimated junction temperature is set as follows: ,in, To predict the junction temperature, Here, P represents the current chip junction temperature, P represents the GPU power consumption, and k represents the coefficient of power consumption on junction temperature. For example, the current chip junction temperature is collected. The temperature was set to 75℃. The GPU power consumption corresponding to the above adjustment parameter combinations was found to be 80W, 70W, and 101W through the performance and power consumption curve model. With k set to 0.05, the estimated junction temperature for the three types of tasks was calculated. The estimated junction temperature for the emergency task was 79℃, for the batch processing task it was 78.5℃, and for the training task it was 80.05℃.

[0087] The temperature safety constraint was then set to 80℃. Comparing the estimated junction temperature of each task with 80℃ revealed that the estimated junction temperature of the training task exceeded this constraint. Therefore, the combination of adjustment parameters needed to be adjusted. First, the voltage was kept constant at 0.98V. Then, the frequency of the training task was reduced from 1.3GHz to 1.25GHz. The GPU power consumption corresponding to the new parameter combination was found to be 95W using the performance-power curve model. The estimated junction temperature was recalculated as 75 + 95 multiplied by 0.05, which equals 79.75℃. This still meets the minimum performance requirements of the task and does not exceed the temperature safety constraint. Subsequently, the parameters of other tasks were verified, and the estimated junction temperatures of the emergency task and batch processing task did not exceed the limit and no adjustment was required.

[0088] The final three sets of adjustment parameters were determined as follows: 1.0V voltage and 1.5GHz frequency for emergency tasks, 0.9V voltage and 1.2GHz frequency for batch processing tasks, and 0.98V voltage and 1.25GHz frequency for training tasks. The GPU was then adjusted synchronously using these parameters to obtain the adjustment results.

[0089] Step 1033: Based on the carbon intensity data and the electricity price information, schedule the calculation tasks to obtain the scheduling results. Based on the utilization efficiency value, perform node allocation processing on the calculation tasks to obtain the node allocation results.

[0090] Step 1033 may specifically include the following steps:

[0091] Step b1: Based on the periodic variation patterns of the carbon intensity data and the electricity price information, identify the distribution characteristics of the target time period to obtain the time period distribution information.

[0092] Among them, the target period refers to the period during which the task is carried out in accordance with the requirements of low carbon intensity and low electricity price, in order to minimize the environmental and economic costs of task execution, including low carbon period and low price period; the period distribution information refers to the time range and periodic distribution pattern data of the target period.

[0093] In step b1, historical carbon intensity data and electricity price information are first collected and their periodic patterns over time are analyzed to identify periods when carbon intensity is consistently low and electricity prices are at their lowest. Then, the time range and recurrence cycle of these periods are extracted to obtain period distribution information.

[0094] For example, historical carbon intensity data and electricity price information can be collected and analyzed to identify target periods with low carbon intensity and low electricity prices each day, thus forming period distribution information.

[0095] Step b2: Based on the time period distribution information, the execution time of the calculation task is arranged to obtain the time arrangement result.

[0096] The time scheduling result refers to the specific execution time plan of the computing task, which can clarify when the task starts and ends. Batch processing tasks are prioritized for execution during low-cost periods, while training tasks are executed during low-carbon periods.

[0097] In step b2, based on the obtained time period distribution information and combined with the characteristics of task type, the batch processing task is fully scheduled to be executed in the low-price period, and the training task is split into multiple segments and concentrated in the low-carbon period to form a time arrangement result with clear execution time of each task; for example, the batch processing task is scheduled to be fully executed in this period, and the training task is split into three segments and concentrated in this period.

[0098] Step b3: Select target data centers from the data center cluster that have a power efficiency value lower than the preset power efficiency threshold and a cooling efficiency higher than the preset cooling efficiency threshold.

[0099] In step b3, preset power efficiency thresholds and cooling efficiency thresholds are first set, and then the power usage efficiency and cooling efficiency of each data center are compared to filter out target data centers whose power usage efficiency is lower than the power efficiency threshold and whose cooling efficiency is higher than the cooling efficiency threshold.

[0100] For example, a preset power efficiency threshold of 1.3 and a cooling efficiency threshold of 85% are set. The data center cluster includes three data centers, A, B, and C, with power efficiency values ​​of 1.2, 1.1, and 1.4, and cooling efficiencies of 88%, 90%, and 82%, respectively. Then, based on the power efficiency threshold and the cooling efficiency threshold, data centers A and B are selected as target data centers.

[0101] Step b4: Based on the load status of the target data center, allocate the corresponding target data center to each computing task to obtain the node allocation result.

[0102] Among them, the load status refers to the real-time collection of core indicators such as CPU utilization, GPU utilization, memory usage, storage I / O busyness and current queue length of tasks to be executed by the hardware monitoring components and resource management system deployed in the target data center. The current resource usage ratio and task load pressure are calculated by summarizing all the core indicators and making a comprehensive judgment.

[0103] The node allocation result refers to the data center allocation scheme corresponding to the computing task, which can clearly identify which data center the task runs in; the target data center refers to a data center that meets the standards for power efficiency and cooling efficiency and has the basic conditions for green operation.

[0104] In step b4, load status data such as the number of tasks and resource utilization rate of each target data center are collected. Then, a suitable target data center is allocated to each task according to the resource requirements of each computing task. At the same time, it is ensured that the load status of each data center is at a balanced level, thereby avoiding excessive load on a single data center, and the node allocation result is obtained.

[0105] For example, if the current load rate of data center A is 40% and that of data center B is 35%, and then execution nodes are allocated according to the resource requirements of the three types of tasks, it is found that urgent tasks with higher resource requirements are allocated to data center B, while batch processing tasks and training tasks are split according to resource requirements and allocated to data centers A and B respectively. After the allocation, the load rate of data center A is 40% + 12% = 52%, and the load rate of data center B is 35% + 18% = 53%. The load balancing threshold is set to a difference of no more than 5%. After calculation, the difference between 52% and 53% is within the threshold, so the nodes are in a balanced state, and the node allocation result is obtained.

[0106] Step 1034: Integrate the time arrangement results and the node allocation results to form an optimized result.

[0107] The optimization result refers to the complete scheduling scheme after integrating the time arrangement result and the node allocation result, which provides the core input for the subsequent construction of decision vectors.

[0108] In step 1034, the obtained time scheduling results and node allocation results are associated and integrated according to the task ID to clarify the execution time period and running data center corresponding to each task, and form a unified optimization result.

[0109] This application adapts energy efficiency strategies according to task type and dynamically adjusts GPU operating parameters, which not only ensures the core execution requirements of different tasks but also avoids the risk of equipment overheating. By combining carbon intensity, electricity price, and data center energy efficiency optimization of task execution time and node arrangement, it can achieve synergistic adaptation between task execution and green and low-cost goals.

[0110] S104. Based on the adjustment results and optimization results, a decision vector is constructed, which is used to represent the scheduling scheme to be determined.

[0111] In one specific implementation, step S104 includes:

[0112] Step 1041: Extract the adjustment parameters corresponding to the GPU operating voltage and frequency from the adjustment results, and extract the node allocation parameters and time scheduling parameters of the computing tasks from the optimization results.

[0113] Among them, the adjustment parameters refer to the core GPU operating parameters extracted from the adjustment results, including the specific values ​​corresponding to the GPU operating voltage and frequency, which are used to clarify the GPU's operating status settings; the node allocation parameters refer to the information related to the execution location of the computing tasks extracted from the optimization results, including the target data center identifier and resource allocation number corresponding to the task, which are used to clarify the running node of the task.

[0114] The time scheduling parameters refer to the computational task execution time-related information extracted from the optimization results. The time scheduling parameters include the task start time, end time, and the division of execution time segments, which are used to clarify the execution sequence of the task.

[0115] In step 1041, the parameter validity is first verified for the adjustment results, and after removing abnormally fluctuating values, the core specific values ​​corresponding to the GPU operating voltage and frequency are selected. At the same time, the unique identifier of the corresponding computing task is associated to ensure that each set of adjustment parameters is accurately matched with the task, and finally a set of adjustment parameters with clear classification is extracted.

[0116] Based on the optimization results, the results are broken down according to two dimensions: execution location and execution time. First, location-related information such as the target data center identifier and resource allocation number is extracted. Then, the rationality is verified in combination with the resource configuration of the data center to confirm that the task requirements match the data center resources and form node allocation parameters.

[0117] Then, extract time-related information such as task start time, end time, and segmented execution time periods, and compare and verify them with the latest completion time constraint of the task to ensure that the time arrangement meets the task requirements and form time arrangement parameters, thereby completing the extraction, verification and classification of three types of core parameters.

[0118] Step 1042: According to the preset encoding rules, combine the adjustment parameters corresponding to the GPU operating frequency and voltage, as well as the node allocation parameters and time arrangement parameters of the computing task, to construct a decision vector.

[0119] The preset encoding rules refer to the pre-defined rules for parameter sorting, format standardization, and numerical mapping to integrate various scattered parameters into a unified vector. These rules ensure the structure and identifiability of the decision vector. For example, the parameters are first sorted in a fixed order: adjustment parameters, node allocation parameters, and time scheduling parameters. The original values ​​of voltage and frequency parameters are retained and labeled with units. The target data center identifier is digitally mapped, and the time parameters are converted to timestamp format. Finally, all standardized parameters are concatenated in order to form a one-dimensional vector decision vector.

[0120] A decision vector is a structured data set formed by integrating adjustment parameters, node allocation parameters, and time scheduling parameters through preset coding rules. It is used to present the pending scheduling scheme intuitively and completely, facilitating subsequent optimization.

[0121] In step 1042, the extracted three types of parameters are first standardized according to the preset encoding rules. That is, the non-numerical center identifier is converted into a digital code through a preset mapping table, the time parameter is converted into a calculable numerical form according to a unified timestamp conversion rule, and the voltage and frequency parameters are supplemented with standard units. Then, according to the logic of "equipment operation priority, execution location second, execution time final", the standardized parameters are arranged in an orderly manner according to a preset fixed order. Finally, all the sorted parameters are concatenated in sequence using a unified separator to form a well-structured and complete one-dimensional structured decision vector, which comprehensively and clearly presents the pending scheduling scheme.

[0122] For example, the parameters related to the three types of computing tasks are first standardized. The non-numerical center identifiers are converted using a preset mapping table, so that data center A is mapped to 1 and data center B is mapped to 2. The time parameters are converted to timestamps according to the unified rule of 3600 seconds per hour. The start time of the emergency task is converted to 32400 for 09:00 and the end time is converted to 36000 for 10:00. The start time of the batch processing task and the training task is converted to 0 for 00:00 and the end time is converted to 21600 for 06:00. The training task is divided into segments, with 02:00 converted to 7200, 02:30 to 9000, 04:30 to 16200, and 05:00 to 18000. The voltage and frequency parameters are supplemented with standard units, namely 1.0V, 1.5GHz, 0.9V, 1.2GHz, 0.98V, and 1.25GHz.

[0123] Then, following the logic of "equipment operation priority, execution location second, execution time finalization", the ranking results of emergency tasks are 1.0V, 1.5GHz, 2, 001, 32400, 36000; batch processing tasks are 0.9V, 1.2GHz, 1, 002, 0, 21600; and training tasks are 0.98V, 1.25GHz, 2, 003, 0, 7200, 9000, 16200, 18000, 21600.

[0124] Finally, using commas as a unified separator, the sorted parameters are concatenated sequentially to obtain the decision vectors for the emergency task: 1.0V, 1.5GHz, 2,001, 32400, 36000; the decision vectors for the batch processing task: 0.9V, 1.2GHz, 1,002, 0, 21600; and the decision vectors for the training task: 0.98V, 1.25GHz, 2,003, 0, 7200, 9000, 16200, 18000, 21600.

[0125] This application extracts and integrates core parameters related to scheduling, transforming scattered adjustment settings, execution locations, and time arrangements into structured decision vectors. This clearly presents pending scheduling schemes, providing a unified and intuitive processing object for subsequent multi-objective optimization and adaptive scheduling, and further ensuring the consistency and efficiency of scheduling decisions.

[0126] S105. Combining the preset multi-objective optimization function, an adaptive scheduling strategy is adopted to perform computing power scheduling on the decision vector to obtain the scheduling result. The multi-objective includes: task completion time, total energy consumption, carbon emissions, and electricity cost.

[0127] In one specific implementation, step S105 includes:

[0128] Step 1051: Determine the completion time constraint based on the latest completion time in the task attribute information; determine the service quality constraint based on the service quality requirements; determine the resource constraint based on the resource requirements in the task attribute information; and determine the temperature safety constraint based on the temperature data.

[0129] Among them, the completion time constraint refers to the time limit determined based on the latest completion time in the task attribute information and the execution time estimated by the task complexity, which is used to ensure that the task is completed within the specified period.

[0130] Service quality constraints refer to performance standards determined based on requirements such as task response speed, data processing accuracy, and operational stability, used to ensure that task execution meets quality standards.

[0131] Resource constraints refer to resource allocation limits determined based on resource requirements in task attribute information, combined with the total amount of hardware resources in the data center cluster and the status of occupied resources, in order to avoid excessive or insufficient resource allocation.

[0132] Temperature safety constraints refer to the maximum operating temperature standards for equipment set based on temperature data, used to prevent equipment from being affected or damaged due to overheating.

[0133] In step 1051, the latest completion time is first extracted from the task attribute information, and the execution time is estimated by combining the complexity indicators such as the amount of computation and data of the task. The earliest start time and the latest start time of the task are determined by subtracting the estimated execution time from the latest completion time, thus forming a completion time constraint. Then, the minimum requirements of core indicators such as response speed, data processing accuracy, and operational stability are determined by referring to the business scenario requirements corresponding to the task, thus forming a service quality constraint.

[0134] Then, based on the resource requirements in the task attribute information and the total amount and proportion of hardware resources such as GPU, CPU, and memory in the data center cluster, the maximum resource allocation and minimum resource guarantee for a single task are determined to form resource constraints. Finally, based on the chip junction temperature safety threshold in the temperature data and combined with the heat dissipation efficiency of the device operation, the maximum allowable temperature for continuous operation of the device is determined to form temperature safety constraints.

[0135] Step 1052: Integrate the completion time constraint, the service quality constraint, the resource constraint, and the temperature safety constraint to obtain a set of constraint conditions.

[0136] Among them, the constraint set refers to the set of constraints formed by integrating completion time constraints, service quality constraints, resource constraints, and temperature safety constraints in a unified format, which is used to provide clear compliance standards for decision vector optimization.

[0137] In step 1052, the determined completion time constraints, service quality constraints, resource constraints, and temperature safety constraints are classified and summarized using a unified constraint description format, and constraint priority identifiers are added, i.e., key constraints are marked as level one and ordinary constraints are marked as level two, thus forming a set of constraints with clear structure and explicit priority.

[0138] Step 1053: Based on task type information and preference labels, the weight coefficients of each objective in the preset multi-objective optimization function are configured differently to obtain an adaptive scheduling strategy.

[0139] Step 1053 may specifically include the following steps:

[0140] Step c1: Determine the number of tasks for each task type based on the task type information.

[0141] In step c1, all received computing tasks are counted according to task type information to determine the number of urgent tasks, batch processing tasks, and training tasks.

[0142] Step c2: Based on the number of tasks and the corresponding preference labels for each task type, a global optimization strategy is used to differentiate the weight coefficients of each objective in the multi-objective optimization function to obtain a weight allocation scheme.

[0143] The weight allocation scheme refers to the configuration scheme formed by differentiating the weight coefficients of each objective in the multi-objective optimization function based on task type, number of tasks, and preference labels. This includes increasing the weight coefficient of task completion time for urgent tasks while decreasing the weight coefficients of total energy consumption, carbon emissions, and electricity costs; increasing the weight coefficients of total energy consumption and electricity costs for batch processing tasks while decreasing the weight coefficients of task completion time and carbon emissions; and increasing the weight coefficient of carbon emissions for training tasks while decreasing the weight coefficient of task completion time.

[0144] In step c2, the analytic hierarchy process (AHP) is used as the global optimization strategy. The preference labels corresponding to task types are used as the criterion layer, and the four objectives of the multi-objective optimization function are used as the scheme layer. The proportion of each task type is combined with the fact that the higher the proportion, the greater the weight of the corresponding preference label. A judgment matrix is ​​constructed and weight coefficients are calculated. The performance priority preference label corresponding to urgent tasks will increase the weight coefficient of task completion time and decrease the weight of other objectives. The energy efficiency priority preference label corresponding to batch processing tasks will increase the weight coefficient of total energy consumption and electricity cost and decrease the weight of other objectives. The segmented execution preference label corresponding to training tasks will increase the weight coefficient of carbon emissions and decrease the weight of other objectives. Finally, the weight allocation scheme is obtained.

[0145] Step c3: Associate the weight allocation scheme with the functions of each objective in the multi-objective optimization function to form an adaptive scheduling strategy.

[0146] Among them, the adaptive scheduling strategy refers to the dynamic scheduling rules formed by associating the weight allocation scheme with the multi-objective optimization function. It can automatically adapt the optimization direction according to the task type and requirements to achieve differentiated balance of multiple objectives.

[0147] In step c3, the target weight coefficients in the weight allocation scheme are substituted into the preset multi-objective optimization function to establish the mapping relationship between the weight coefficients and each objective function, thereby forming an adaptive scheduling strategy that can automatically adapt the optimization direction according to the task type.

[0148] Step 1054: Based on the set of constraints and the adaptive scheduling strategy, optimize the decision vector to obtain multiple candidate decision vectors.

[0149] Among them, candidate decision vectors refer to multiple structured data sets that meet the constraints and have better multi-objective optimization effects after multiple rounds of optimization processing of the initial decision vectors, and are used for subsequent selection of the optimal solution.

[0150] In step 1054, a genetic algorithm is used to optimize the decision vector. The constructed decision vector is used as the initial population, the constraint set is used as the compliance judgment standard, and the multi-objective optimization function corresponding to the adaptive scheduling strategy is used as the fitness function. Then, when initializing the population, the population size is set to 50, each individual corresponds to a decision vector, and the top 30% of high-quality individuals are selected based on fitness values ​​by the selection operator. Then, the parameter fragments of the high-quality individuals are randomly exchanged and recombined by the crossover operator, and the parameters of the recombined individuals are slightly randomly adjusted by the mutation operator to generate a new generation of population.

[0151] Subsequently, after each generation of the population is generated, it is first checked whether it meets the constraint set, and individuals that violate the first-level constraints are removed, individuals that violate the second-level constraints are corrected, and then the fitness value of the remaining individuals is calculated. The higher the fitness value, the better the multi-objective optimization effect. Finally, the iteration is repeated many times until the average fitness value of the population fluctuates by no more than 5% for 5 consecutive generations, and all high-quality compliant individuals retained in the iteration process are used as multiple candidate decision vectors.

[0152] Step 1055: Select the target decision vector from all the candidate decision vectors using the Pareto optimization method, and convert the target decision vector into the corresponding scheduling result.

[0153] The target decision vector refers to the optimal structured data set that comprehensively meets the multi-objective optimization requirements selected from all candidate decision vectors through the Pareto optimization method. It is the core basis for forming the final scheduling result. The scheduling result refers to the specific executable instructions converted from the target decision vector. The scheduling result includes GPU running parameter setting instructions, task execution location allocation instructions, task execution timing arrangement instructions, etc., which are used to directly guide the implementation of computing power scheduling.

[0154] In step 1055, the specific values ​​of the four objectives corresponding to each candidate decision vector—task completion time, total energy consumption, carbon emissions, and electricity cost—are first calculated. Then, the Pareto optimization method is used for non-dominated sorting. For any two candidate decision vectors A and B, if all objectives of A are not inferior to B and at least one objective is superior to B, then B is dominated and eliminated. The remaining non-dominated individuals form the non-dominated solution set. Then, the crowding degree of each candidate decision vector in the non-dominated solution set is calculated. The crowding degree is determined by the sum of the distances between the solution and its neighboring solutions in each objective dimension. The larger the distance, the higher the crowding degree, which means that the solution is more scarce in the solution set.

[0155] Subsequently, based on the core business requirements, the solution with the highest congestion and that conforms to the core preferences of the task type is selected from the non-dominated solution set as the target decision vector. Finally, the adjustment parameters in the target decision vector are converted into GPU running parameter setting instructions, the node allocation parameters are converted into task execution location allocation instructions, the time scheduling parameters are converted into task execution timing scheduling instructions, and all instructions are integrated to form the final scheduling result.

[0156] This application integrates multi-dimensional constraints and adopts an adaptive scheduling strategy, combined with optimization algorithms and Pareto optimal selection, to achieve a multi-objective balance between task completion time, total energy consumption, carbon emissions and electricity costs. This ensures that the scheduling results meet the preset compliance requirements and can adapt to the core needs of different tasks, thereby improving the scientificity and practicality of computing power scheduling.

[0157] Figure 3 This is a schematic diagram illustrating a specific implementation of a green computing power scheduling system based on multi-objective optimization, as provided in this application. (Refer to...) Figure 3 The system may include:

[0158] The acquisition module 31 is used to receive multiple computing tasks and acquire task attribute information corresponding to each computing task, wherein the task attribute information includes task type information.

[0159] The acquisition module 32 is used to acquire GPU power and temperature data of each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data of each time period in the area covered by each data center from the application programming interface connected to the power grid.

[0160] The adjustment module 33 is used to dynamically adjust the GPU power of each data center based on the task type information and the temperature data to obtain the adjustment result. Based on the carbon intensity data, the utilization efficiency value, and the electricity price information, it optimizes the scheduling of each computing task to obtain the optimization result.

[0161] The construction module 34 is used to construct a decision vector based on the adjustment results and optimization results, the decision vector being used to represent the pending scheduling scheme.

[0162] The scheduling module 35 is used to combine a preset multi-objective optimization function and adopt an adaptive scheduling strategy to perform computing power scheduling on the decision vector to obtain the scheduling result. The multi-objectives include: task completion time, total energy consumption, carbon emissions and electricity cost.

[0163] The green computing power scheduling system based on multi-objective optimization in this application is used to implement the aforementioned green computing power scheduling method based on multi-objective optimization. Therefore, the specific implementation of the green computing power scheduling system based on multi-objective optimization can be found in the embodiment section of the green computing power scheduling method based on multi-objective optimization above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0164] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described green computing power scheduling methods based on multi-objective optimization.

[0165] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described green computing power scheduling methods based on multi-objective optimization.

[0166] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0167] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the green computing power scheduling method based on multi-objective optimization described above.

[0168] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0169] The foregoing has provided a detailed description of a green computing power scheduling system and method based on multi-objective optimization provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A green computing power scheduling method based on multi-objective optimization, characterized in that, include: Receive multiple computing tasks and obtain task attribute information corresponding to each computing task, wherein the task attribute information includes task type information; Collect GPU power and temperature data from each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data for each time period in the area covered by each data center from the application programming interface connected to the power grid. Based on the task type information and the temperature data, the GPU power of each data center is dynamically adjusted to obtain the adjustment result. Based on the carbon intensity data, the utilization efficiency value, and the electricity price information, the scheduling of each computing task is optimized to obtain the optimization result. Based on the adjustment and optimization results, a decision vector is constructed, which is used to represent the scheduling scheme to be determined. Combining a preset multi-objective optimization function, an adaptive scheduling strategy is adopted to perform computing power scheduling on the decision vector to obtain the scheduling result. The multi-objectives include: task completion time, total energy consumption, carbon emissions, and electricity cost. The step involves combining a preset multi-objective optimization function with an adaptive scheduling strategy to perform computational power scheduling on the decision vector, obtaining scheduling results including: The following constraints are determined based on the latest completion time in the task attribute information, the service quality constraints based on the service quality requirements, the resource constraints based on the resource requirements in the task attribute information, and the temperature safety constraints based on the temperature data. The completion time constraint is a time limit determined based on the latest completion time in the task attribute information and the estimated execution time based on the task complexity. The service quality constraint is a performance standard determined based on the task's response speed, data processing accuracy, and operational stability requirements. The resource constraint is a resource allocation limit determined based on the resource requirements in the task attribute information and the total amount of hardware resources in the data center cluster and the occupied resources. The temperature safety constraint is a maximum operating temperature standard for the equipment set based on the temperature data. The completion time constraint, the service quality constraint, the resource constraint, and the temperature safety constraint are integrated to obtain a set of constraint conditions; Based on task type information and preference labels, the weight coefficients of each objective in the preset multi-objective optimization function are configured differently to obtain an adaptive scheduling strategy. Based on the set of constraints and the adaptive scheduling strategy, the decision vector is optimized to obtain multiple candidate decision vectors; The Pareto optimization method is used to select the target decision vector from all the candidate decision vectors, and the target decision vector is converted into the corresponding scheduling result; The construction of the decision vector based on the adjustment and optimization results includes: Extract the adjustment parameters corresponding to the GPU operating voltage and frequency from the adjustment results, and extract the node allocation parameters and time scheduling parameters of the computing tasks from the optimization results; According to the preset encoding rules, the adjustment parameters corresponding to the GPU operating frequency and voltage, as well as the node allocation parameters and time scheduling parameters of the computing task, are combined to construct a decision vector. The preset encoding rules are to sort the adjustment parameters, node allocation parameters, and time scheduling parameters in a fixed order, retain the original values ​​of the voltage and frequency parameters, use digital mapping for the target data center identifier, and convert the time scheduling parameters into a timestamp format.

2. The method of claim 1, wherein, The process involves dynamically adjusting the GPU power of each data center based on the task type information and the temperature data to obtain adjustment results. Then, based on the carbon intensity data, the utilization efficiency value, and the electricity price information, the scheduling of each computing task is optimized to obtain optimization results, including: Based on the preference tags corresponding to the task type information, configure the corresponding energy efficiency strategy for the computation task to obtain the strategy allocation result. The task type information is an emergency task, a batch processing task, or a training task. Based on the strategy allocation results and the pre-established performance and power consumption curve model, combined with the temperature data, the GPU operating voltage and frequency are coordinated and adjusted to obtain an adjustment result that meets temperature safety constraints. Based on the carbon intensity data and the electricity price information, the calculation tasks are scheduled in time to obtain the scheduling results. Based on the utilization efficiency value, the calculation tasks are assigned nodes to obtain the node assignment results. The time scheduling results and the node allocation results are integrated to form an optimized result.

3. The method according to claim 2, characterized in that, The method, based on the strategy allocation results and a pre-established performance and power consumption curve model, combined with the temperature data, performs coordinated adjustment of the GPU operating voltage and frequency to obtain an adjustment result that meets temperature safety constraints, including: Based on the strategy allocation results and the pre-established performance and power consumption curve model, determine the combination of adjustment parameters for GPU operating voltage and frequency; Based on the temperature data and preset temperature safety constraints, the combination of adjustment parameters is constrained, and the operating voltage and frequency of the GPU are synchronously adjusted using the constrained combination of adjustment parameters to obtain the adjustment result.

4. The method according to claim 2, characterized in that, The process of scheduling computational tasks based on the carbon intensity data and the electricity price information to obtain scheduling results, and allocating nodes for computational tasks based on the utilization efficiency value to obtain node allocation results, includes: Based on the periodic variation patterns of the carbon intensity data and the electricity price information, the distribution characteristics of the target time period are identified to obtain the time period distribution information; Based on the time period distribution information, the execution time of the computing tasks is arranged to obtain the time arrangement result; Select target data centers from the data center cluster that have a power efficiency value lower than a preset power efficiency threshold and a cooling efficiency higher than a preset cooling efficiency threshold. Based on the load status of the target data center, a corresponding target data center is assigned to each computing task to obtain the node allocation result.

5. The method according to claim 1, characterized in that, The adaptive scheduling strategy is obtained by differentiating the weight coefficients of each objective in the preset multi-objective optimization function based on task type information and preference labels, including: Based on task type information, determine the number of tasks for each task type; Based on the number of tasks and corresponding preference labels for each task type, a global optimization strategy is used to differentiate the weight coefficients of each objective in the multi-objective optimization function to obtain a weight allocation scheme. The weight allocation scheme is associated with the functions of each objective in the multi-objective optimization function to form an adaptive scheduling strategy.

6. A green computing power scheduling system based on multi-objective optimization, characterized in that, To implement the green computing power scheduling method based on multi-objective optimization as described in claim 1, the method includes: The acquisition module is used to receive multiple computing tasks and acquire task attribute information corresponding to each computing task, wherein the task attribute information includes task type information. The data acquisition module is used to collect GPU power and temperature data from each data center in the data center cluster, calculate the power efficiency value of each data center, and obtain electricity price information and carbon intensity data for each time period in the area covered by each data center from the application programming interface connected to the power grid. The adjustment module is used to dynamically adjust the GPU power of each data center based on the task type information and the temperature data to obtain the adjustment result; and to optimize the scheduling of each computing task based on the carbon intensity data, the utilization efficiency value, and the electricity price information to obtain the optimization result. A construction module is used to construct a decision vector based on the adjustment and optimization results, wherein the decision vector is used to represent the pending scheduling scheme; The scheduling module is used to combine a preset multi-objective optimization function and adopt an adaptive scheduling strategy to perform computing power scheduling on the decision vector to obtain the scheduling result. The multi-objectives include: task completion time, total energy consumption, carbon emissions, and electricity cost.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the green computing power scheduling method based on multi-objective optimization as described in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the green computing power scheduling method based on multi-objective optimization as described in any one of claims 1 to 5.

Images