Calculation energy cooperation business planning method and device and related equipment

Abstract

The invention provides a computing energy collaborative service planning method and device and related equipment, and is applied to the technical field of computers. Comprising the following steps: acquiring equipment configuration information, service configuration information and electricity price information; based on the equipment configuration information and the service configuration information, predicting the equipment performance information and the service performance information by using a time sequence prediction model, and outputting multi-index prediction results corresponding to the equipment performance information and the service performance information respectively; determining a correlation between GPU power and GPU load of each device based on the device configuration information; constructing a corresponding objective function and constraint conditions of the objective function according to the multi-index prediction results corresponding to the equipment performance information and the business performance information, the correlation between the GPU power and the GPU load of each piece of equipment and the electricity price information; and under the constraint of the constraint condition, solving the target function to obtain a business scheduling strategy when the operation cost is minimum. Resource waste can be avoided, and operation cost can be reduced.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a computing power collaborative business planning method, apparatus and related equipment. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent computing centers, as core infrastructure, are being built at an accelerated pace. Compared with traditional data centers, intelligent computing centers have significantly higher equipment power density, leading to a surge in annual power consumption and a significant increase in operating costs.

[0003] To address this challenge, the industry is dynamically scheduling latency-insensitive services, concentrating their execution during green electricity periods; simultaneously charging energy storage devices when green electricity supply is sufficient, and utilizing these devices to supply power during peak electricity price periods. This solution aims to increase the proportion of green electricity consumption and reduce enterprise operating costs and carbon emissions. However, current service scheduling strategies suffer from resource waste.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This disclosure provides a computing power collaborative business planning method, apparatus and related equipment, which generates an optimal business scheduling strategy by comprehensively considering computing efficiency and economic benefit indicators, effectively avoiding resource waste and significantly reducing operating costs.

[0006] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0007] According to one aspect of this disclosure, a computing power collaborative service planning method is provided. The method includes: acquiring equipment configuration information, service configuration information, and electricity price information; based on the equipment configuration information and the service configuration information, using a time-series prediction model to predict equipment performance information and service performance information, and outputting multi-index prediction results corresponding to each of the equipment performance information and the service performance information; determining the correlation between GPU power and GPU load for each device based on the equipment configuration information; constructing a corresponding objective function and constraints for the objective function with the goal of minimizing operating costs, based on the multi-index prediction results corresponding to each of the equipment performance information and the service performance information, the correlation between GPU power and GPU load for each device, and the electricity price information; and solving the objective function under the constraints to obtain the service scheduling strategy with the minimum operating cost.

[0008] In some embodiments, the step of predicting device performance information and service performance information using a time-series prediction model based on the device configuration information and the service configuration information, and outputting multi-indicator prediction results corresponding to each of the device performance information and the service performance information, includes: modeling multiple time-series indicators of device performance information and service performance information, wherein the time-series indicators include CPU core count, memory resource usage, video memory resource usage, and GPU utilization; and predicting a preset time period based on a preset time granularity using a time-series prediction model based on the device configuration information and the service configuration information, and outputting multi-indicator prediction results corresponding to each of the device performance information and the service performance information, wherein the multi-indicator prediction results corresponding to the device performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each device, and the multi-indicator prediction results corresponding to the service performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each service.

[0009] In some embodiments, the device configuration information includes at least the GPU model. The step of determining the relationship between GPU power and GPU load for each device based on the device configuration information includes: collecting power data of each GPU model under different loads to form a sample dataset; and using a multinomial regression algorithm to fit the sample dataset to obtain the GPU power and GPU load relationship model and model parameters for each GPU model of each device.

[0010] In some embodiments, the step of constructing a corresponding objective function based on the multi-indicator prediction results corresponding to the device performance information and the service performance information, the correlation between GPU power and GPU load of each device, and the electricity price information, with the goal of minimizing operating costs, includes: dividing the preset time period into green electricity periods and non-green electricity periods based on the electricity price information; constructing a total operating cost objective function, wherein the total operating cost is the sum of the cost of green electricity periods and the cost of non-green electricity periods; wherein the cost of green electricity periods and the cost of non-green electricity periods are calculated based on the multi-indicator prediction results corresponding to the device performance information and the service performance information, the correlation between GPU power and GPU load of each device, and the electricity price of the corresponding time period.

[0011] In some embodiments, during non-green electricity periods, the GPU utilization of each device is equal to the predicted GPU utilization of each device minus the sum of the GPU utilization of all services migrated to the green electricity period on each device; during green electricity periods, the GPU utilization of each device is equal to the GPU utilization of each device plus the sum of the GPU utilization of all services migrated to the green electricity period on each device.

[0012] In some embodiments, when a service uses multiple GPU cards from the same device, the GPU utilization rate of the service is the average of the GPU utilization rates of the GPU cards used by the service.

[0013] In some embodiments, the constraints include: at a preset time, the number of CPU cores of each device is greater than the sum of the predicted CPU core usage of each device and the predicted CPU core usage of services on each device; at a preset time, the memory resources of each device are greater than the sum of the predicted memory resource usage of each device and the predicted memory resource usage of services on each device; at a preset time, the video memory resources of each device are greater than the sum of the predicted video memory resource usage of each device and the predicted video memory resource usage of services on each device.

[0014] In some embodiments, the method further includes setting resource allocation thresholds for various resources of each device.

[0015] In some embodiments, the method further includes: allowing resource preemption during service migration; after a resource is preempted, the service whose resource was preempted may be restarted when the resource is released, or may be scheduled to be executed on other devices.

[0016] According to another aspect of this disclosure, a computing power collaborative service planning device is also provided. The device includes: an acquisition module for acquiring equipment configuration information, service configuration information, and electricity price information; a prediction module for predicting equipment performance information and service performance information based on the equipment configuration information and the service configuration information using a time-series prediction model, and outputting multi-index prediction results corresponding to each of the equipment performance information and the service performance information; a determination module for determining the correlation between GPU power and GPU load of each device based on the equipment configuration information; a construction module for constructing a corresponding objective function and constraints of the objective function with the goal of minimizing operating costs, based on the multi-index prediction results corresponding to the equipment performance information and the service performance information, the correlation between GPU power and GPU load of each device, and the electricity price information; and a solution module for solving the objective function under the constraints of the constraints to obtain the service scheduling strategy with the minimum operating cost.

[0017] According to another aspect of this disclosure, an electronic device is also provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the computing power collaborative service planning method described in any one of the preceding claims by executing the executable instructions.

[0018] According to another aspect of this disclosure, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the computing power collaborative service planning method described in any one of the preceding claims.

[0019] According to another aspect of this disclosure, a computer program product is also provided, comprising: a computer program or instructions that, when executed by a processor, implement the computing power collaborative service planning method of any one of the above.

[0020] An embodiment of this disclosure provides a computing power collaborative service planning method, comprising: acquiring device configuration information, service configuration information, and electricity price information; based on the device configuration information and service configuration information, using a time-series prediction model to predict device performance information and service performance information, and outputting multi-index prediction results corresponding to each of the device performance information and service performance information; determining the correlation between GPU power and GPU load for each device based on the device configuration information; constructing a corresponding objective function and constraints for the objective function with the goal of minimizing operating costs, based on the multi-index prediction results corresponding to each of the device performance information and service performance information, the correlation between GPU power and GPU load for each device, and electricity price information; and solving the objective function under the constraints to obtain the service scheduling strategy with the minimum operating cost. This embodiment uses a time-series prediction model to predict multiple indicators of equipment performance and service performance information, providing an accurate data foundation for service scheduling and avoiding unreasonable scheduling decisions due to information lag or estimation errors. It also determines the correlation between GPU power and load based on equipment configuration information, breaking the limitations of traditional linear assumptions and making equipment power calculations more closely reflect actual operating conditions, providing an accurate basis for energy cost optimization. Finally, it constructs an objective function by combining the multi-indicator prediction results, GPU power-load relationship, and electricity price information, and obtains the cost-optimal service scheduling strategy through solving under constraints. This effectively balances service operation needs with power distribution system energy cost control, achieving the goal of minimizing power distribution system operating costs.

[0021] Furthermore, by deeply coupling equipment characteristics, business characteristics, and electricity price differences into scheduling decisions, this method can not only accurately match the business operation needs and equipment resource supply at different times, but also reduce the proportion of high-energy-consuming business operations during non-green electricity periods through scientific scheduling, thereby indirectly reducing the overall carbon emissions of the intelligent computing system. It takes into account both cost-effectiveness and green operation requirements, providing the most optimized business scheduling strategy for the refined operation of intelligent computing centers, effectively avoiding resource waste and significantly reducing operating costs.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0024] Figure 1 This diagram illustrates a system architecture of a computing power collaborative service planning method according to an embodiment of the present disclosure. Figure 2 This diagram illustrates a flowchart of a computing power collaborative service planning method according to an embodiment of the present disclosure; Figure 3 This invention discloses a flowchart of a method for predicting multi-indicator prediction results according to an embodiment of the present disclosure. Figure 4 This invention discloses a flowchart illustrating a method for determining the relationship between GPU power and GPU load in various devices according to an embodiment of the present disclosure. Figure 5 This diagram illustrates a method for constructing an objective function according to an embodiment of the present disclosure. Figure 6 A flowchart illustrating a specific computing power collaborative business planning method according to an embodiment of this disclosure is shown. Figure 7 This diagram illustrates a computing power collaborative service planning device according to an embodiment of the present disclosure. Figure 8 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0025] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0026] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] As mentioned earlier, compared to traditional data centers, intelligent computing centers have significantly higher power density, leading to a surge in annual power consumption. For example, a traditional data center consumes approximately 5-10kW, while a single rack in an intelligent computing center can reach 20-50kW, and a single large intelligent computing center (such as 1000 PFlops) can consume hundreds of millions of kilowatt-hours annually. In response to the high energy consumption of intelligent computing centers, various regions have proposed green energy-saving requirements in their construction plans. Coordinated optimization of "computing power-power (computing-power)" has become a key path to balance computing power demand and energy constraints, with dynamic computing power scheduling being the fundamental condition for achieving this coordination. Current mainstream computing-power coordination optimization solutions include: centrally scheduling latency-insensitive services to operate during green electricity periods, charging energy storage devices during periods of abundant green electricity, and activating energy storage power supply during peak electricity price periods. This model can both increase the proportion of green electricity consumption and simultaneously reduce enterprise operating costs and carbon emission levels.

[0028] Existing solutions assume a linear relationship between equipment load and energy consumption. However, in actual testing, the relationship is a curve with a gradually decreasing slope. Therefore, when the same service is scheduled to run on devices with different loads, the additional energy consumption will be different. Existing solutions do not take this difference into account. Furthermore, when generating scheduling policies, existing solutions are based on the amount of resources requested by the service and do not take into account the actual resource usage of the service. This may result in some latency-insensitive services not being able to be satisfied during green electricity periods, causing some resource waste.

[0029] In view of this, the embodiments of this disclosure rely on time-series prediction models to predict multiple indicators of equipment performance information and service performance information, providing an accurate data foundation for service scheduling and avoiding unreasonable scheduling decisions due to information lag or estimation errors. Furthermore, based on equipment configuration information, the correlation between GPU power and load is determined, breaking the limitations of traditional linear assumptions and making equipment power calculations more closely reflect actual operating conditions, providing an accurate basis for energy cost optimization. Finally, an objective function is constructed by combining the multi-indicator prediction results, GPU power-load relationship, and electricity price information, and the cost-optimal service scheduling strategy is obtained through solving under constraints. This effectively balances service operation needs with power distribution system energy cost control, achieving the goal of minimizing power distribution system operating costs. Moreover, by deeply coupling equipment characteristics, service characteristics, and electricity price differences into scheduling decisions, this method can not only accurately match service operation needs and equipment resource supply at different times, but also reduce the proportion of high-energy-consuming service operations during non-green electricity periods through scientific scheduling, indirectly reducing the overall carbon emissions of the intelligent computing system. It balances cost-effectiveness and green operation requirements, providing an optimized service scheduling strategy for the refined operation of intelligent computing centers, effectively avoiding resource waste and significantly reducing operating costs.

[0030] To facilitate understanding, before introducing the embodiments of this disclosure, the following explanations are provided for several terms involved in the embodiments of this disclosure: To facilitate understanding of the embodiments of this disclosure, the key terms involved in the embodiments will be explained below: Computational efficiency refers to the ratio of effective computing power to resource consumption, used to quantify the balance between device computing efficiency and energy consumption costs. In this method, computational efficiency is specifically defined as the ratio of device (such as GPU, NPU, etc.) utilization rate to its operating power. The higher the ratio, the stronger the effective computing power that the device can provide per unit of energy consumption.

[0031] Peak electricity: refers to the period when the power grid is at its peak load. During this period, the overall demand for electricity in society is strong, and the power grid faces greater pressure. In order to regulate the balance of power supply and demand, a higher electricity price is usually applied.

[0032] Off-peak electricity: refers to the period when the power grid is at its lowest point in terms of electricity load. During this period, the overall demand for electricity in society is low, the power supply resources of the power grid are relatively abundant, and in order to encourage the consumption of electricity, a lower electricity price is usually applied.

[0033] Green electricity, also known as green power, specifically refers to electricity generated using renewable energy technologies such as wind, solar, hydro, and biomass energy. Its core characteristic is that the power generation process produces almost no greenhouse gas emissions, has minimal impact on the ecological environment, falls under the category of clean energy, and typically enjoys a cost advantage in electricity pricing policies, resulting in relatively low electricity prices.

[0034] A Graphics Processing Unit (GPU) is a hardware device specifically designed for handling graphics rendering and parallel computing tasks. Initially used primarily in computer graphics (such as game rendering and image editing), it has, with technological advancements and its powerful parallel computing capabilities, been widely adopted in intelligent computing scenarios (such as deep learning model training and inference, and large-scale data processing). It can efficiently handle multi-threaded parallel tasks and improve the processing efficiency of complex computing scenarios.

[0035] The central processing unit (CPU) is the core computing unit of a computer system. It is responsible for executing the instructions of computer programs, coordinating the scheduling of various hardware resources (such as memory and peripherals) in the system, and handling core operations such as logical judgment, data processing, and task control.

[0036] Random Access Memory (RAM) is a high-speed storage device in a computer system used for temporary storage of data and program instructions. It is characterized by fast read and write speeds, but data is lost when power is off. Its main function is to provide the CPU with the data and instructions needed for real-time computation, reducing the frequency of the CPU directly accessing low-speed storage devices (such as hard drives), thus improving overall computational efficiency. It is the core space for temporary data storage during the operation of business programs.

[0037] Video Random Access Memory (VRAM): Also known as video random access memory, it is a high-speed storage device dedicated to GPUs, specifically used to store data required for GPU computation (such as graphics textures, model parameters, and intermediate calculation results in deep learning). Its performance (such as capacity and bandwidth) directly affects the GPU's computing efficiency, avoiding latency caused by frequent data exchange between the GPU and memory. This ensures that the GPU efficiently processes data in graphics rendering and parallel computing (such as AI tasks), making it a key storage resource for ensuring the full potential of the GPU.

[0038] The Prophet time series forecasting model is a time series forecasting algorithm designed to predict future trends, seasonality, and holiday effects by decomposing time series and fitting them with machine learning techniques.

[0039] The specific implementation methods of the embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0040] Figure 1 An exemplary application system architecture diagram is shown, to which the computing power collaborative service planning method of the embodiments of this disclosure can be applied. For example... Figure 1 As shown, the system architecture may include terminal device 101, network 102 and server 103.

[0041] Network 102 is a medium used to provide a communication link between terminal device 101 and server 103, and can be a wired network or a wireless network.

[0042] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to Local Area Networks (LANs), Metropolitan Area Networks (MANs), Wide Area Networks (WANs), mobile, wired or wireless networks, private networks, or any combination of virtual private networks. In some embodiments, technologies and / or formats, including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Networks (VPNs), and Internet Protocol Security (IPSec) can be used to encrypt all or some links. In other embodiments, custom and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0043] Terminal device 101 can be various electronic devices, including but not limited to smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, wearable devices, augmented reality devices, virtual reality devices, etc.

[0044] Optionally, the client of the application installed on different terminal devices 101 may be the same, or the client of the same type of application based on different operating systems. Depending on the terminal platform, the specific form of the application client may also be different; for example, the application client may be a mobile client, a PC client, etc.

[0045] Server 103 can be a server that provides various services, such as a backend management server that supports the device operated by the user using terminal device 101. The backend management server can analyze and process received requests and other data, and feed the processing results back to the terminal device.

[0046] Optionally, the server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0047] Those skilled in the art will know that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative; any number of terminal devices, networks, and servers can be included depending on actual needs. This disclosure does not limit the scope of the embodiments.

[0048] Under the above system architecture, this disclosure provides a computing power collaborative service planning method, which can be executed by any electronic device with computing processing capabilities.

[0049] In some embodiments, the computing power collaborative service planning method provided in this disclosure can be executed by the terminal device of the above-described system architecture; in other embodiments, the computing power collaborative service planning method provided in this disclosure can be executed by the server in the above-described system architecture; in still other embodiments, the computing power collaborative service planning method provided in this disclosure can be implemented by the terminal device and the server in the above-described system architecture through interaction.

[0050] Figure 2 A flowchart illustrating a computing power collaborative service planning method according to an embodiment of this disclosure is shown. Figure 2 As shown in the embodiments of this disclosure, the computing power collaborative service planning method includes the following steps: S202, obtain equipment configuration information, service configuration information, and electricity price information.

[0051] In this embodiment, the device configuration information includes the device's hardware parameters and resource specifications. This includes, but is not limited to, device identifier, rated power, number of CPU cores, memory capacity, disk capacity (not required for compute-in-store devices), number of GPU cards, single-card memory capacity, and rated memory bandwidth. Among them, the device identifier ensures that each device has a unique identity, so that subsequent business can be accurately associated with the device; rated power refers to the maximum power consumption of the device when it is running at full load, which is used for subsequent energy consumption calculation and power supply capacity matching; CPU core count records the number of physical cores of the CPU installed in the device; memory capacity records the total physical memory capacity that the device can provide; disk capacity is only collected for non-compute-storage devices, recording the total capacity of the hard disk installed locally in the device. For compute-storage devices, since storage resources are provided by an independent storage cluster, this field value can be marked as "-"; GPU card count records the number of independent GPU accelerator cards installed in the device, and GPU card model information can also be distinguished as an extended field; single card memory capacity is for devices equipped with GPU cards, recording the total memory capacity of a single GPU card. If the device is equipped with multiple models of GPU cards, the memory capacity of different models and their corresponding quantities can be recorded in a preset format; rated memory bandwidth refers to the data transfer rate of a single GPU card's memory.

[0052] The service configuration information includes the resource requirements, execution rules, and priority settings for intelligent computing services, used to clarify the service's resource requirements and time constraints. This includes, but is not limited to, service identifier, service type, service priority, device identifier, service execution time, CPU resource allocation, memory resource allocation, GPU card allocation, and disk resource allocation. The service identifier is a unique identifier for the service; the service type can be categorized into three types based on its operating mode: "online task," "planned offline task," and "offline task triggered by a preceding task." "Planned offline tasks" are the focus of subsequent resource scheduling optimization and should be marked as "highly important" during information collection; the service priority indicates the level of importance of the service, used for prioritizing scheduling in case of resource conflicts; the device identifier records the device identifier initially planned or already allocated for the service. If the service is in a pending allocation state, this field value is "pending allocation" or another field value indicating a pending allocation state; the service execution time is recorded differently according to the service type. Planned offline tasks record the specific scheduled execution time, while offline tasks triggered by preceding tasks have uncertain execution times, so this field value can be... The value of this field is assumed to be "-1". Since online tasks need to run continuously for 24 hours, this field value can be "0:00-24:00". The CPU resource allocation records the number of CPU cores requested by the service, and must be less than or equal to the number of CPU cores on the device where the service is located. The memory resource allocation records the memory capacity requested by the service, and must be less than or equal to the memory capacity on the device where the service is located. The GPU card allocation is only for GPU-intensive service collection and records the number of GPU cards requested by the service, and must be less than or equal to the number of GPU cards on the device where the service is located. The disk resource allocation is for services in non-disk-compute separation scenarios and records the local disk capacity requested by the service. In storage-compute separation scenarios, since storage resources are allocated from an independent storage cluster and recorded separately in the storage configuration information, this field value can be empty.

[0053] Electricity price information refers to the electricity cost parameters of the intelligent computing center, used for energy cost calculation and green electricity priority dispatch in subsequent business operations. The data is collected from the electricity price policy provided by the power grid company and the green electricity procurement agreement of the intelligent computing center. Peak electricity price records the unit price of electricity used during peak periods of the power grid. The peak period time division is implemented according to the power grid company's standards and can be read from the power grid price module of the energy management subsystem. The peak electricity price can be the daily peak period division and price data updated at 0:00 every day. Green electricity price records the unit price of green electricity (such as wind power and photovoltaic). The green electricity price can be adjusted according to the quarterly green electricity procurement agreement and can be read from the green electricity procurement management module. The effective date is updated synchronously when adjusting. Green electricity time period records the available time period for green electricity access by the intelligent computing center.

[0054] In some embodiments, data validity verification is required through an information verification module: Device configuration information must ensure unique device identifiers and that hardware parameters match the device model; service configuration information must ensure that resource allocation does not exceed the device's maximum capacity and that the execution time format is compliant; electricity price information must ensure that the price is positive and that time periods do not overlap. After successful verification, the above information is stored in the system's unified information database. This database can employ a distributed storage architecture, supporting concurrent reading from multiple subsystems. A data update log is also established to record the time, operator, and data changes for each information collection and update, ensuring data traceability and providing accurate and complete foundational data for subsequent business scheduling decisions.

[0055] S204, based on equipment configuration information and service configuration information, uses a time-series prediction model to predict equipment performance information and service performance information, and outputs the prediction results of multiple indicators corresponding to each of the equipment performance information and service performance information.

[0056] In this embodiment, the time-series prediction model can be the Prophet time-series prediction model, and independent sub-models are constructed for different performance indicators to ensure the accuracy of the prediction results. The prediction process can be carried out with a time granularity of 5 minutes to predict the performance information of devices and services in the intelligent computing center within a preset time period (such as the next 24 hours).

[0057] Specifically, in actual prediction, forecasts are performed from two dimensions: device and business. Device-level forecasts focus on individual devices, and the final output includes a unique device identifier, along with 5-minute granular forecasts of metrics such as real-time power consumption, CPU core count, memory usage, disk usage, GPU utilization, video memory utilization, and real-time video memory bandwidth. Business-level forecasts focus on individual services, and the output includes a unique service identifier, along with 5-minute granular forecasts of metrics such as CPU core count, memory usage, disk usage, GPU utilization, and video memory utilization. It should be noted that when the intelligent computing center includes multiple devices and multiple services, it is necessary to predict the multi-metric forecasts for all devices and all services within the intelligent computing center.

[0058] S206 determines the relationship between GPU power and GPU load for each device based on device configuration information.

[0059] It should be noted that related solutions typically simplify the relationship between equipment load and energy consumption to a linear one. However, actual testing has revealed that the true relationship is a curve with a gradually decreasing slope. This means that the additional energy consumption generated when the same service is scheduled to run on devices under different load conditions will vary, thus affecting the energy consumption optimization effect of resource scheduling. In the business scheduling scenario of intelligent computing centers, GPU devices, as the core computing power components, account for the highest proportion of the total energy consumption of the devices, directly impacting the energy utilization efficiency of the intelligent computing center.

[0060] In view of this, after predicting the device performance information and service performance information in this embodiment, it is necessary to clarify the relationship between the GPU power and GPU load of each device based on the device configuration information, so as to provide an accurate basis for subsequent service scheduling strategies and solve the problem of energy consumption calculation deviation caused by linear assumptions in the prior art.

[0061] As we can understand, power refers to the electrical energy consumed by a GPU device per unit of time, used to characterize the rate of energy consumption; energy consumption refers to the total electrical energy consumed by a GPU device within a preset time period. Power and energy consumption satisfy the formula: Energy Consumption = Power × Running Time. When the running time is fixed, power directly determines the total energy consumption. GPU load is the actual workload undertaken by the GPU device, expressed as a percentage from 0% to 100%, reflecting the workload intensity; GPU utilization is the ratio of the actual computing time of the GPU to the total running time, also expressed as a percentage, measuring usage efficiency. When utilization increases, the proportion of time the GPU is in computing state increases, and the load will also increase accordingly; the two are positively correlated. That is, without any restrictions, due to the GPU device's own energy-saving strategy, when the GPU utilization is low, the GPU will operate with relatively low power to maintain basic operation and response speed; while when utilization increases, the GPU's power will increase significantly to meet higher computing demands. Therefore, power is most strongly correlated with GPU utilization. In power-load curve fitting, since load and utilization are positively correlated, the relationship between power and load can be indirectly reflected by fitting the relationship between power and GPU utilization.

[0062] It's important to note that device configuration information is crucial for differentiating the hardware characteristics of various devices. This includes parameters such as GPU card model, rated power per card, and number of GPU cards. Different GPU card models employ different hardware architectures, including manufacturing processes, core counts, and thermal designs, resulting in variations in energy efficiency. The relationship between GPU power and load for each device includes the power-load relationship for multiple GPU models. The device's power is calculated by adding the base power to the sum of the power of all GPU cards on the device, as shown in the following formula:

[0063] in, This represents the power of the d-th device in the intelligent computing center. This represents the base power of the d-th device. This represents the sum of the power of all GPUs on the d-th device. This represents the power of the k-th GPU on the d-th device, where K represents the total number of GPUs on the d-th device.

[0064] It should be noted that other components on the device, such as the CPU, also generate additional power consumption when there is business activity, but this is much smaller than the power consumption of the GPU. Therefore, this impact is ignored in this implementation. In actual applications, other power consumption may also be added, and this embodiment does not impose any restrictions on this.

[0065] Existing solutions assume a linear relationship between equipment load and energy consumption. However, actual tests show that this relationship is a curve with a decreasing slope, leading to differences in additional energy consumption when the same service is scheduled to devices with different loads. Existing solutions do not consider this difference. This embodiment determines the correlation between GPU power and load for each device by fitting the relationship between GPU power and utilization, avoiding the linear assumption. This makes the device power calculation more closely reflect actual operating conditions, providing an accurate basis for energy cost optimization, effectively avoiding resource waste, and reducing operating costs.

[0066] S208: Based on the multi-indicator prediction results corresponding to the equipment performance information and business performance information, the correlation between GPU power and GPU load of each device, and electricity price information, construct the corresponding objective function and the constraints of the objective function with the goal of minimizing operating costs.

[0067] This embodiment revolves around the economic objective function of integrated computing and energy consumption. This objective function aims to comprehensively consider the relationship between the computing power usage and energy consumption of the equipment, calculate the total operating cost of the equipment in 24-hour (day) increments, and provide a quantitative basis for the output of daily business scheduling optimization strategies.

[0068] Since the time granularity is preset to 5 minutes, a 24-hour day is divided into 288 time periods, with a total number of time periods of n=288 and a duration of T=5 minutes for each time period. Assuming that the equipment load and power remain stable within a single time period, the energy consumption of that time period can be directly calculated based on the power. Let the total number of devices be D. The computing power-energy consumption relationship of the d-th device (d=1, 2, ..., D) follows the fitting logic described above, and its parameters... , , It is tied to the device's GPU model and hardware characteristics.

[0069] The energy efficiency objective function of the computing-energy integration system takes the total cost C as the optimization objective. The total cost is obtained by summing the energy consumption costs of all devices at all times throughout the day, and the formula is defined as follows:

[0070] Where n represents the total number of time periods, and D represents the total number of devices. Let i be the electricity price during the i-th time period. Let be the energy consumption of the d-th device during the i-th time period.

[0071]

[0072] in, This represents the power of the d-th device in the intelligent computing center during the i-th time period. The power remains constant within a time period by default. Indicates the duration of each time period. This represents the base power of the d-th device in the i-th time period. This represents the sum of the power of all GPUs on the d-th device during the i-th time period. This represents the sum of the power of all GPUs on the d-th device during the i-th time period. This represents the power of the k-th GPU on the d-th device during the i-th time period. This represents the total number of GPUs on the d-th device. , , The parameters represent the computing power-energy consumption fitting function of the d-th device. This represents the utilization rate of the k-th GPU in the i-th time period for the d-th device.

[0073] S210, under the constraints, solve the objective function to obtain the service scheduling strategy that minimizes operating costs.

[0074] In this embodiment, since service scheduling affects the GPU utilization of devices, the predicted GPU utilization value for each device cannot be used directly. Instead, the actual load and power need to be calculated based on the predicted service volume and device performance. It should be noted that due to the spatiotemporal scheduling characteristics of services, the optimization problem is a complex mixed-integer programming problem. Specifically, optimization solvers and other solution models can be used to find a set of scheduling schemes that minimize the total cost C while satisfying all constraints. The output is the device scheduling strategy of the power distribution system for each time period, including: the power allocation of each device in each time period, the load allocation of each device in each time period, the service scheduling strategy, and the total energy consumption and cost for each time period.

[0075] It should be noted that when a service is migrated to a specific device, the CPU, memory, and video memory resource limits of that device must be met. Specifically, at a predetermined time, the number of CPU cores on each device must be greater than the sum of the predicted CPU core usage for each device and the predicted CPU core usage for the service on each device; at a predetermined time, the memory resources on each device must be greater than the sum of the predicted memory resource usage for each device and the predicted memory resource usage for the service on each device; and at a predetermined time, the video memory resources on each device must be greater than the sum of the predicted video memory resource usage for each device and the predicted video memory resource usage for the service on each device. It should be noted that the predicted CPU core usage for services on each device is specifically the sum of the predicted CPU core usage for services planned to be migrated to the d-th device at time t; similarly, the predicted memory resource usage for services on each device is specifically the sum of the predicted memory resource usage for services planned to be migrated to the d-th device at time t; and the predicted video memory resource usage for services on each device is specifically the sum of the predicted video memory resource usage for services to be migrated to the d-th device at time t; where time t is the preset time.

[0076] In some embodiments, the limiting conditions are expressed by the following formula:

[0077] in, This represents the number of CPU cores in the d-th device. This represents the predicted CPU core usage of the d-th device at time t. This represents the sum of predicted CPU core usage values ​​for services planned to be migrated to the d-th device at time t. This represents the predicted CPU core usage of the s-th service planned to be migrated to the d-th device at time t. This represents the memory resources of the d-th device. This represents the predicted memory resource usage of the d-th device at time t. This represents the sum of the predicted memory resource usage of services planned to be migrated to the d-th device at time t. This represents the predicted memory resource usage of the s-th service, which is planned to be migrated to the d-th device, at time t. This represents the video memory resources of the d-th device. This represents the predicted memory usage of the d-th device at time t. This represents the sum of predicted GPU memory usage at time t for services to be migrated to the d-th device. This represents the predicted memory usage of the s-th service planned to be migrated to the d-th device at time t.

[0078] In this embodiment, a time-series prediction model is used to predict multiple indicators of equipment performance and service performance information, providing an accurate data foundation for service scheduling and avoiding unreasonable scheduling decisions due to information lag or estimation errors. Furthermore, the correlation between GPU power and load is determined based on equipment configuration information, breaking the limitations of traditional linear assumptions and making equipment power calculations more closely reflect actual operating conditions, providing an accurate basis for energy cost optimization. Finally, an objective function is constructed by combining the multi-indicator prediction results, GPU power-load relationship, and electricity price information. The optimal service scheduling strategy is obtained by solving the problem under constraints, effectively balancing service operation needs with power distribution system energy cost control, and achieving the goal of minimizing power distribution system operating costs.

[0079] Furthermore, by deeply coupling equipment characteristics, business characteristics, and electricity price differences into scheduling decisions, this method can not only accurately match the business operation needs and equipment resource supply at different times, but also reduce the proportion of high-energy-consuming business operations during non-green electricity periods through scientific scheduling, thereby indirectly reducing the overall carbon emissions of the intelligent computing system. It takes into account both cost-effectiveness and green operation requirements, providing the most optimized business scheduling strategy for the refined operation of intelligent computing centers, effectively avoiding resource waste and significantly reducing operating costs.

[0080] Figure 3 This invention discloses a flowchart illustrating a method for predicting multi-indicator prediction results according to an embodiment of the present invention. Figure 3 As shown in the embodiments of this disclosure, the method for predicting device performance information and service performance information using a time-series prediction model based on device configuration information and service configuration information, and outputting multi-index prediction results corresponding to each of the device performance information and service performance information, includes the following steps: S302 models multiple time-series indicators of device performance information and service performance information, including CPU core count, memory resource usage, video memory resource usage, and GPU utilization.

[0081] In this embodiment, Prophet time-series prediction models are constructed for four indicators: CPU core count, memory resource usage, video memory resource usage, and GPU utilization.

[0082] It should be noted that the modeling logic for each indicator needs to be designed in accordance with its data characteristics. For example, for CPU core count and memory resource usage, since they are greatly affected by the business execution stage, a "business stage" should be introduced as a custom regression variable in the Prophet model. By using the business execution time and business type obtained through step S202, the business stage corresponding to each time point in the historical data can be labeled to improve the model's fit to the resource fluctuation pattern. For video memory resource usage and GPU utilization, since the resource consumption of GPU-intensive businesses has periodic peak characteristics, the periodic component can be strengthened in the model. By utilizing the Prophet model's ability to capture periodic data, the fluctuation pattern of video memory and GPU resources can be accurately fitted.

[0083] In some embodiments, the modeling process can also set up a model evaluation and iteration mechanism, using average absolute error and root mean square error as evaluation indicators. If the evaluation indicator of a certain indicator model exceeds the preset percentage of the historical average of the indicator, the model parameters need to be adjusted retrospectively and retrained based on historical data until the model error meets the preset threshold, ensuring that the four types of time series indicator models have high prediction accuracy.

[0084] S304, based on device configuration information and service configuration information, uses a time-series prediction model to predict a preset time period based on a preset time granularity, and outputs multi-indicator prediction results corresponding to device performance information and service performance information respectively. Among them, the multi-indicator prediction results corresponding to device performance information include the predicted values ​​of CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each device. The multi-indicator prediction results corresponding to service performance information include the predicted values ​​of CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each service.

[0085] In this embodiment, the preset time granularity refers to the smallest unit of time to be divided when making predictions, which can be set to "5 minutes". The preset time period refers to the future time range to be predicted, which can be set to "the next 24 hours". Specifically, the prediction process needs to be executed along two paths: one along the device dimension and the other along the business dimension.

[0086] Equipment performance information prediction: Taking a single device as the prediction object, the device configuration information obtained in step S202, such as the number of CPU cores, memory capacity, and number of GPU cards, is used as a resource upper limit constraint. This information is input into four types of time-series prediction models for the corresponding device: CPU core count model, memory resource usage model, video memory resource usage model, and GPU utilization model. The model is based on the business plan associated with the device within the next 24 hours, obtained by matching the "device identifier of the business" with the "business execution time" extracted from the business configuration information in step S202. Combined with historical resource usage patterns, the prediction results are output according to a preset time granularity. The predicted value for the number of CPU cores used is... This indicates that the predicted value of device memory resource usage is used... This indicates that the predicted value of device memory resource usage is used... This indicates that the predicted value of device GPU utilization is used... This indicates that 'd' represents the device identifier, 't' represents the predicted time point, and the time interval T is 5 minutes. It should be noted that all predicted values ​​must not exceed the hardware resource limits specified in the device configuration information. ≤Number of CPU cores in the device (≤100%) If the predicted value exceeds the upper limit, the model will automatically correct the output result based on hardware constraints. It is understood that the above example only uses a single device as the prediction object; in actual applications, the relevant device performance information for all devices needs to be preset.

[0087] Business performance information prediction: Taking a single business as the prediction object, the business configuration information obtained in step S202, such as business type, resource allocation, and execution time, is used as business feature input. Four time-series prediction models are applied to the corresponding business: CPU core count model, memory resource usage model, video memory resource usage model, and GPU utilization model. The model combines historical resource consumption data from the business's execution history and outputs prediction results for the business's execution period within the next 24 hours at a 5-minute granularity. The predicted value for the business's CPU core usage is... This indicates that the predicted value of business memory resource usage is used... This indicates that the predicted value of the service's video memory resource usage is used... This indicates that the predicted value of business GPU utilization is used... This indicates that 's' is the service identifier and 't' is the prediction time point. It should be noted that the prediction result can be linked to the prediction result of the device where the service resides in step S202. For example, if a service 's' is assigned to device 'd', then... The sum needs to be combined with Matching is used to avoid discrepancies in resource usage statistics.

[0088] In some embodiments, in addition to the CPU, memory, video memory resource usage and GPU utilization that are the focus of this embodiment, the following other performance and resource indicators may also be included, which are not limited in this embodiment.

[0089] Because related technologies generate scheduling strategies based on the amount of resources requested by services without considering the actual resource usage of services, some latency-insensitive services cannot be satisfied during green electricity periods, resulting in some resource waste. In this embodiment, a time-series prediction model is used to output the predicted value of the actual resource usage of services. During scheduling, resources can be allocated based on the predicted value rather than the requested amount, thereby improving the utilization rate of green electricity and the service satisfaction rate, and reducing resource waste.

[0090] In addition, the scheduling strategies of related technologies lack data support, resulting in a mismatch between resource allocation and actual needs. The multi-indicator prediction results of devices and services output in this embodiment provide a clear quantitative basis for subsequent scheduling strategies. Some non-urgent GPU services can be scheduled to other devices with lower GPU utilization in advance to avoid device overload leading to increased service latency. This ensures that the scheduling strategy is feasible and reasonable, and improves the overall operating efficiency of the intelligent computing center.

[0091] In some embodiments, the device configuration information includes at least the GPU model. Figure 4 This disclosure illustrates a flowchart of a method for determining the relationship between GPU power and GPU load in various devices, as shown in the embodiments of the present disclosure. Figure 4 As shown in the embodiments of this disclosure, the method for determining the relationship between GPU power and GPU load of each device based on device configuration information includes the following steps: S402 collects power data for each GPU model under different loads to form a sample dataset.

[0092] In this embodiment, the model of the GPU installed in each device is used as the classification dimension. For each model of GPU device, the device performance monitoring module collects data points, that is, records the real-time load value and the corresponding real-time power value or utilization rate of the GPU at fixed time intervals. The collected GPU model, load value, power value data and GPU utilization rate are classified and organized by model. After removing outliers, a sample dataset specific to each GPU model is formed, providing raw data support for the subsequent construction of power-load relationship curves.

[0093] S404 uses a multinomial regression algorithm to fit the sample dataset, obtaining the GPU power and GPU load relationship model and model parameters for each GPU model of each device.

[0094] In this embodiment, the multinomial regression algorithm can accurately fit the nonlinear curve relationship between GPU power and load, and is more in line with the actual data characteristics than linear regression.

[0095] Taking model A as an example, fitting is performed based on the sample dataset collected in step S402. Model construction is based on GPU utilization. To characterize the load, a quadratic polynomial is used as the basic form. The relationship between the power and utilization of GPU model A is shown in the following formula:

[0096] in, For GPU power of model A, For GPU utilization of model A, , , These are the model parameters to be solved.

[0097] During the fitting process, taking model A as an example, the effective sample data of model A's GPU collected by S402 is substituted into the multinomial regression algorithm. The sum of squared residuals between the predicted power and the actual power is minimized using the least squares method, and the parameters are finally obtained: , , The corresponding model expression is .

[0098] It should be noted that the use of a quadratic polynomial as the basic form is only for illustrative purposes. In actual applications, the polynomial regression algorithm can also adopt other forms, as long as they conform to the relationship between GPU power and GPU load corresponding to the GPU model. This embodiment does not impose any restrictions on this.

[0099] This embodiment determines the correlation between GPU power and load for each device by fitting the relationship between GPU power and GPU load, avoiding linear assumptions and making device power calculations more consistent with actual operating conditions. This provides an accurate basis for energy consumption cost optimization, effectively avoids resource waste, and reduces operating costs.

[0100] In some embodiments, for better optimization analysis, the objective function can be split into two parts. Among them, using... This represents the costs incurred during non-green electricity periods, while This represents the cost during the green electricity period. The calculation logic for these two parts is essentially the same as the original objective function; however, in practical applications, it is calculated separately for different time periods (non-green electricity periods and green electricity periods). The total number of time periods corresponding to non-green electricity periods and green electricity periods are n1 and n2, respectively.

[0101] Figure 5 This diagram illustrates a method flowchart for constructing an objective function according to an embodiment of the present disclosure, such as... Figure 5As shown in the embodiments of this disclosure, the objective function constructed based on the multi-indicator prediction results corresponding to device performance information and service performance information, the correlation between GPU power and GPU load of each device, and electricity price information, with the goal of minimizing operating costs, includes the following steps: S502 divides a preset time period into green electricity periods and non-green electricity periods based on electricity price information.

[0102] S504, construct the objective function for total operating cost, which is the sum of the cost during green electricity periods and the cost during non-green electricity periods; wherein, the cost during green electricity periods and the cost during non-green electricity periods are calculated based on the multi-indicator prediction results corresponding to equipment performance information and business performance information, the correlation between GPU power and GPU load of each device, and the electricity price of the corresponding period.

[0103] In this embodiment, the objective function is expressed by the following formula:

[0104] in, This represents the total cost over a preset time period. This represents the cost incurred during non-green electricity periods within a preset time frame, while This indicates the cost of green electricity during the preset time period.

[0105] In some embodiments, during non-green electricity periods, the GPU utilization of each device is equal to the predicted GPU utilization of each device minus the sum of the GPU utilization of all services migrated to the green electricity period on each device; during green electricity periods, the GPU utilization of each device is equal to the GPU utilization of each device plus the sum of the GPU utilization of all services migrated to the green electricity period on each device.

[0106] Please refer to the following formula for details:

[0107] Where n1 represents the total number of non-green electricity periods. This indicates the time period outside of green electricity usage periods, where D represents the total number of devices. This represents the power of the d-th device at time t. This indicates that the d-th device at time t is... One GPU utilization prediction value, This represents the sum of the service power to be migrated in or out on the d-th device at time t. Let m represent the predicted GPU utilization of the s-th service at time t, m represent the total number of services to be migrated in or out of the d-th device, and T represent the duration of each time period. Let n1 be the electricity price in the i-th time period within the non-green electricity period, and n2 represent the total number of time periods corresponding to the green electricity period. The time period indicating green electricity periods Let n be the electricity price during the i-th time period outside of the green electricity period, and n1 and n2 sum to n. It should be noted that due to business relocation, This represents the predicted GPU utilization of the d-th device at time t when no migrations in or out occur.

[0108] In some embodiments, when a service uses multiple GPU cards from the same device, the service’s GPU utilization rate is the average of the GPU utilization rates of the GPU cards used by the service.

[0109] In some embodiments, in order to ensure the service quality of the original latency-sensitive services after the service migration, the method further includes setting resource allocation thresholds for various resources of each device.

[0110] In this embodiment, a resource allocation threshold r is set during migration. Taking device CPU resources as an example, the amount of allocable resources is... r, during the solution process, replace the CPU constraint with The same applies to other resource indicators, namely... , This ensures the quality of business operations even in the event of a minor resource disruption. It should be noted that the threshold values ​​for different resources can be the same or different; this embodiment does not impose any restrictions on this.

[0111] In some embodiments, the method further includes: allowing resource preemption during service migration; after a resource is preempted, the service whose resource was preempted may be restarted when the resource is released, or may be scheduled to be executed on other devices.

[0112] In this embodiment, since the scheduled services are all latency-insensitive services and are allowed to be rerun after failure, when a latency-sensitive service has a sudden demand for resources, the resources of the latency-insensitive service can be preempted, and the task can be rerun when the resources are released again, or it can be scheduled to be executed on other devices.

[0113] Figure 6 This illustration shows a flowchart of a specific computing power collaborative business planning method according to an embodiment of the present disclosure, such as... Figure 6 As shown in the embodiments of this disclosure, the computing power collaborative service planning method includes the following steps: S601 obtains the service configuration, load, classification, cabinet power consumption, and energy measurement of green electricity periods, green electricity prices, and peak-valley electricity prices from the intelligent computing center side.

[0114] In this embodiment, the data processing module can be used to process static and time-series data from the intelligent computing center and the power supply side for subsequent model building and strategy generation. This includes equipment rated resource data, operating load data, energy consumption data, business static resource allocation data, business classification and priority data, green electricity periods, green electricity prices, and peak-valley electricity prices.

[0115] S602, Build a forecasting model for equipment and business resource requirements.

[0116] In this embodiment, the resource prediction module can be used to use a time-series prediction model (such as Prophet) to predict various resource utilization indicators of equipment and services for the next 24 hours.

[0117] S603, Fitting of equipment power load relationship.

[0118] In this embodiment, the power-load relationship fitting module can be used to perform GPU load and power sampling on different models of devices, and fit the load-power curves of different devices based on the sampling data.

[0119] S604 constructs an integrated computing and energy objective function based on business information, resource demand forecasts, non-green electricity periods, green electricity periods, and power compliance curves, with the goal of minimizing operating costs.

[0120] In this embodiment, economic cost is used as the direct optimization objective, and the relationship between energy consumption and computing power is optimized. Compared with existing solutions, it does not assume that the energy consumption remains constant when the task is executed at different times and on different devices. Therefore, the objective function is an integrated computing and energy-saving economic objective function that combines computing efficiency.

[0121] S605, based on constraints, generates the business migration strategy that is optimal for the objective function.

[0122] In this embodiment, based on the objective function solving module, the predicted data of equipment and service resource requirements and the power load curve are substituted into the objective function. Based on the rated resource generation constraints, the approximate optimal solution of the objective function is obtained through a genetic algorithm.

[0123] S606, Implement security measures for business operations.

[0124] Since scheduling is based on actual resource demand, resource over-allocation may occur. In this embodiment, a business operation security assurance module is used to formulate and execute a security operation strategy to ensure the safe operation of the business itself and other businesses before and after the business migration.

[0125] Based on the same inventive concept, this disclosure also provides a computing power collaborative service planning device, as described in the following embodiments. Since the principle by which this device solves the problem is similar to that of the above-described method embodiments, the implementation of this device embodiment can refer to the implementation of the above-described method embodiments, and repeated details will not be elaborated further.

[0126] Figure 7 This diagram illustrates a computing power collaborative service planning device according to an embodiment of the present disclosure, such as... Figure 7 As shown, the device includes: an acquisition module 71, used to acquire device configuration information, service configuration information, and electricity price information; a prediction module 72, used to predict device performance information and service performance information based on the device configuration information and the service configuration information using a time-series prediction model, and output the multi-index prediction results corresponding to the device performance information and the service performance information respectively; a determination module 73, used to determine the correlation between GPU power and GPU load of each device based on the device configuration information; a construction module 74, used to construct a corresponding objective function and constraints of the objective function with the goal of minimizing operating cost, based on the multi-index prediction results corresponding to the device performance information and the service performance information, the correlation between GPU power and GPU load of each device, and the electricity price information; and a solution module 75, used to solve the objective function under the constraints of the constraints to obtain the service scheduling strategy with the minimum operating cost.

[0127] In some embodiments, the prediction module is specifically used to: model multiple time-series indicators of device performance information and service performance information, the time-series indicators including CPU core count, memory resource usage, video memory resource usage, and GPU utilization; based on the device configuration information and the service configuration information, use the time-series prediction model to predict a preset time period based on a preset time granularity, and output the multi-indicator prediction results corresponding to the device performance information and the service performance information, wherein the multi-indicator prediction results corresponding to the device performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each device, and the multi-indicator prediction results corresponding to the service performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each service.

[0128] In some embodiments, the device configuration information includes at least the GPU model. The determining module is specifically used to: collect power data of each GPU model under different loads to form a sample dataset; and use a multinomial regression algorithm to fit the sample dataset to obtain the GPU power and GPU load relationship model and model parameters corresponding to each GPU model of each device.

[0129] In some embodiments, the construction module is specifically used to: divide the preset time period into green electricity periods and non-green electricity periods based on electricity price information; construct a total operating cost objective function, wherein the total operating cost is the sum of the cost of green electricity periods and the cost of non-green electricity periods; wherein the cost of green electricity periods and the cost of non-green electricity periods are calculated based on the multi-indicator prediction results corresponding to the equipment performance information and the business performance information, the correlation between the GPU power and GPU load of each device, and the electricity price of the corresponding time period.

[0130] In some embodiments, during non-green electricity periods, the GPU utilization of each device is equal to the predicted GPU utilization of each device minus the sum of the GPU utilization of all services migrated to the green electricity period on each device; during green electricity periods, the GPU utilization of each device is equal to the GPU utilization of each device plus the sum of the GPU utilization of all services migrated to the green electricity period on each device.

[0131] In some embodiments, when a service uses multiple GPU cards from the same device, the GPU utilization rate of the service is the average of the GPU utilization rates of the GPU cards used by the service.

[0132] In some embodiments, the constraints include: at a preset time, the number of CPU cores of each device is greater than the sum of the predicted CPU core usage of each device and the predicted CPU core usage of services on each device; at a preset time, the memory resources of each device are greater than the sum of the predicted memory resource usage of each device and the predicted memory resource usage of services on each device; at a preset time, the video memory resources of each device are greater than the sum of the predicted video memory resource usage of each device and the predicted video memory resource usage of services on each device.

[0133] In some embodiments, the apparatus further includes a service assurance module, configured to set resource allocation thresholds for various resources of each device.

[0134] In some embodiments, the service assurance module is specifically used to: allow resource preemption during service migration; after a resource is preempted, the service whose resource was preempted will be restarted when the resource is released, or it will be scheduled to be executed on other devices.

[0135] It should be noted that the examples and application scenarios implemented by the modules in the above device embodiments and the corresponding steps in the method embodiments are the same, but are not limited to the content disclosed in the above method embodiments. It should also be noted that the above modules, as part of the device, can be executed in a computer system such as a set of computer-executable instructions.

[0136] Those skilled in the art will understand that various aspects of this disclosure can be implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which can be collectively referred to herein as a "circuit", "module" or "system".

[0137] Based on the same inventive concept, this disclosure also provides an electronic device, which includes: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the computing power collaborative service planning method described above by executing the executable instructions. Since the principle by which this electronic device embodiment solves the problem is similar to that of the above method embodiment, the implementation of this electronic device embodiment can refer to the implementation of the above method embodiment, and repeated details will not be described again.

[0138] The following reference Figure 8 To describe an electronic device 800 according to such an embodiment of the present disclosure. Figure 8 The electronic device 800 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0139] like Figure 8 As shown, the electronic device 800 is manifested in the form of a general-purpose computing device. The components of the electronic device 800 may include, but are not limited to: at least one processing unit 810, at least one storage unit 820, and a bus 830 connecting different system components (including storage unit 820 and processing unit 810).

[0140] The storage unit stores program code that can be executed by the processing unit 810, causing the processing unit 810 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 810 can perform the following steps of the above method embodiments: obtaining device configuration information, service configuration information, and electricity price information; based on the device configuration information and service configuration information, using a time-series prediction model to predict device performance information and service performance information, and outputting multi-index prediction results corresponding to each of the device performance information and service performance information; determining the correlation between GPU power and GPU load for each device based on the device configuration information; based on the multi-index prediction results corresponding to each of the device performance information and service performance information, the correlation between GPU power and GPU load for each device, and the electricity price information, constructing a corresponding objective function and constraints for the objective function with the goal of minimizing operating costs; and solving the objective function under the constraints to obtain the service scheduling strategy with the minimum operating cost.

[0141] Storage unit 820 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 8201 and / or cache memory 8202, and may further include a read-only memory (ROM) 8203.

[0142] The storage unit 820 may also include a program / utility 8204 having a set (at least one) of program modules 8205, such program modules 8205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0143] Bus 830 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0144] Electronic device 800 can also communicate with one or more external devices 840 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 800, and / or with any device that enables electronic device 800 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 850. Furthermore, electronic device 800 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 860. As shown, network adapter 860 communicates with other modules of electronic device 800 via bus 830. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 800, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0145] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0146] Based on the same inventive concept, this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the computing power collaborative service planning method described above. Since the principle by which this computer-readable storage medium embodiment solves the problem is similar to that of the above method embodiments, the implementation of this computer-readable storage medium embodiment can refer to the implementation of the above method embodiments, and repeated details will not be elaborated further.

[0147] More specific examples of computer-readable storage media in this disclosure may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0148] In this disclosure, a computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting a program for use by or in connection with an instruction execution system, apparatus, or device.

[0149] Optionally, the program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0150] In practical implementation, program code for performing the operations of this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0151] Based on the same inventive concept, this disclosure also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the computing power collaborative service planning method of any one of the above method embodiments. Since the principle by which this computer program product embodiment solves the problem is similar to that of the above method embodiments, the implementation of this computer program product embodiment can refer to the implementation of the above method embodiments, and repeated details will not be elaborated further.

[0152] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0153] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0154] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0155] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A computing power collaborative service planning method, characterized in that, The method includes: Obtain equipment configuration information, service configuration information, and electricity price information; Based on the device configuration information and the service configuration information, a time-series prediction model is used to predict the device performance information and the service performance information, and output the multi-indicator prediction results corresponding to the device performance information and the service performance information respectively. Based on the device configuration information, the relationship between GPU power and GPU load for each device is determined. Based on the multi-indicator prediction results corresponding to the device performance information and the service performance information, the correlation between the GPU power and GPU load of each device, and the electricity price information, a corresponding objective function and the constraints of the objective function are constructed with the goal of minimizing operating costs. Under the constraints described, the objective function is solved to obtain the service scheduling strategy that minimizes operating costs.

2. The computing power collaborative service planning method according to claim 1, characterized in that, The step of predicting device performance information and service performance information using a time-series prediction model based on the device configuration information and service configuration information, and outputting multi-index prediction results corresponding to each of the device performance information and service performance information, includes: Modeling is performed on multiple time-series indicators of device performance information and service performance information, including CPU core count, memory resource usage, video memory resource usage, and GPU utilization. Based on the device configuration information and the service configuration information, a time-series prediction model is used to predict a preset time period based on a preset time granularity, and outputs multi-indicator prediction results corresponding to the device performance information and the service performance information, respectively. The multi-indicator prediction results corresponding to the device performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each device. The multi-indicator prediction results corresponding to the service performance information include predicted values ​​for CPU core usage, memory resource usage, video memory resource usage, and GPU utilization for each service.

3. The computing power collaborative service planning method according to claim 1, characterized in that, The device configuration information includes at least the GPU model, and determining the relationship between GPU power and GPU load for each device based on the device configuration information includes: Power data of each GPU model under different loads were collected to form a sample dataset. The sample dataset was fitted using a multinomial regression algorithm to obtain the relationship model between GPU power and GPU load for each GPU model of each device, as well as the model parameters.

4. The computing power collaborative service planning method according to claim 2, characterized in that, Based on the multi-index prediction results corresponding to the device performance information and the service performance information, the correlation between the GPU power and GPU load of each device, and the electricity price information, a corresponding objective function is constructed with the goal of minimizing operating costs, including: Based on electricity price information, the preset time period is divided into green electricity periods and non-green electricity periods; A total operating cost objective function is constructed, wherein the total operating cost is the sum of the cost during green electricity periods and the cost during non-green electricity periods; wherein the cost during green electricity periods and the cost during non-green electricity periods are calculated based on the multi-indicator prediction results corresponding to the equipment performance information and the business performance information, the correlation between the GPU power and GPU load of each device, and the electricity price of the corresponding period.

5. The computing power collaborative service planning method according to claim 4, characterized in that, During non-green electricity periods, the GPU utilization of each device is equal to the predicted GPU utilization of each device minus the sum of the GPU utilization of all services migrated to the green electricity period on each device; During green electricity periods, the GPU utilization of each device is equal to the sum of the GPU utilization of the individual device and the GPU utilization of all services migrated to the individual device.

6. The computing power collaborative service planning method according to claim 4, characterized in that, When a service uses multiple GPU cards from the same device, the GPU utilization rate of the service is the average of the GPU utilization rates of each GPU card used by the service.

7. The computing power collaborative service planning method according to any one of claims 2 to 6, characterized in that, The constraints include: At a predetermined time, the number of CPU cores on each device is greater than the sum of the predicted CPU core usage for each device and the predicted CPU core usage for the services on each device. At a predetermined time, the memory resources of each device are greater than the sum of the predicted memory resource usage of each device and the predicted memory resource usage of the services on each device. At a predetermined time, the video memory resources of each device are greater than the sum of the predicted video memory resource usage of each device and the predicted video memory resource usage of the services on each device.

8. The computing power collaborative service planning method according to claim 7, characterized in that, The method further includes: Set resource allocation thresholds for various resources of each device.

9. The computing power collaborative service planning method according to claim 1, characterized in that, Also includes: Resource preemption is allowed during business migration; After resources are preempted, the services whose resources were preempted will resume operation when the resources are released, or they will be scheduled to be executed on other devices.

10. A computing power collaborative service planning device, characterized in that, The device includes: The acquisition module is used to acquire device configuration information, service configuration information, and electricity price information. The prediction module is used to predict device performance information and service performance information based on the device configuration information and the service configuration information using a time-series prediction model, and output the multi-indicator prediction results corresponding to the device performance information and the service performance information respectively. The determination module is used to determine the relationship between GPU power and GPU load for each device based on the device configuration information; The construction module is used to construct a corresponding objective function and the constraints of the objective function based on the multi-index prediction results corresponding to the device performance information and the service performance information, the correlation between the GPU power and GPU load of each device, and the electricity price information, with the goal of minimizing the operating cost. The solution module is used to solve the objective function under the constraints of the given conditions to obtain the service scheduling strategy that minimizes the operating cost.

11. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the computing power collaborative service planning method according to any one of claims 1 to 9 by executing the executable instructions.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the computing power collaborative service planning method according to any one of claims 1 to 9.

13. A computer program product comprising: A computer program or instruction, characterized in that, when executed by a processor, the computer program or instruction implements the computing power collaborative service planning method according to any one of claims 1 to 9.

Images