Multi-model time-sharing computing power metering method and device of intelligent computing cloud platform

Abstract

The application provides a multi-model time-sharing computing power metering method and device of an intelligent computing cloud platform, and relates to the technical fields of intelligent computing centers, intelligent computing centers, computing power infrastructure and intelligent computing cloud technology.The multi-model time-sharing computing power metering method of the intelligent computing cloud platform comprises the following steps: S1, acquiring CPU utilization rate, GPU index data, storage index data, network index data, business index data, environment and energy efficiency index data; S2, generating a load curve through an LSTM model, outputting a time period label through a random forest classification model, and generating index weights through an XGBoost regression model; S3, outputting a first computing power value through a linear regression model; S4, outputting an energy efficiency factor through a decision tree model, outputting a business factor through a K-Means clustering model, and outputting a load factor based on the load curve and actual load data; and S5, calculating a second computing power value based on the load factor, the business factor, the energy efficiency factor and the first computing power value.The application can solve the problem of high difficulty in computing power metering.

Description

Technical Field

[0001] This invention relates to the fields of intelligent computing centers, smart computing centers, computing infrastructure, and smart cloud technologies, specifically to a multi-model time-sharing computing power metering method and device for an intelligent computing cloud platform. Background Technology

[0002] With the rapid development of artificial intelligence technology, "intelligent computing centers" and "smart computing centers" have emerged.

[0003] An "intelligent computing center" refers to a facility that provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models) by utilizing large-scale heterogeneous computing resources, including general-purpose and intelligent computing power. Intelligent computing centers encompass facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0004] "Intelligent computing center" includes, but is not limited to, "intelligent computing center".

[0005] "Intelligent computing center" or artificial intelligence computing center is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting artificial intelligence computing architecture.

[0006] "Computing power" is the core of "intelligent computing center" and "smart computing center". It is the ability of computer equipment or computing / data center to process parameters. It is the ability of computer hardware and software to work together to execute a certain computing requirement. It is the computing power to achieve the target result output by processing parameter data. It is a new type of productivity that integrates parameter computing power, network carrying capacity and data storage capacity. It mainly provides services to society through computing power infrastructure.

[0007] With the explosive growth of artificial intelligence, big data models, scientific computing, and the Internet of Things, computing power has risen from a technical support role to a new type of production factor (on par with land, labor, capital, and data). For any production factor to achieve efficient allocation, fair trade, and sustainable development, a scientific, reliable, and operable measurement system must be established. Measuring computing power is far more complex and challenging than measuring traditional resources such as electricity, data traffic, and storage. The fundamental reason is that computing power is not a "homogeneous, linearly superimposed" physical quantity, but a "dynamic capability" highly dependent on task type, system architecture, software stack, and context. It cannot be "loaded into a meter" like electricity, nor can it be "counted by bytes" like data traffic. Therefore, building a reliable, fair, and tradable computing power measurement system has become a core underlying challenge in developing "computing power networks," "East-West data computing," and "AI infrastructure."

[0008] It is evident that since the emergence of intelligent computing centers, measuring computing power has been extremely difficult, and there is an urgent need for an effective and scientific method for measuring computing power. This problem has always been a pressing issue to be solved in this field. Summary of the Invention

[0009] This invention provides a multi-model time-sharing computing power measurement method and device for intelligent computing cloud platforms, in order to solve the problem that computing power measurement is difficult in the existing technology and that there is an urgent need for an effective and scientific computing power measurement method.

[0010] To solve the above problems, the present invention is implemented as follows: In a first aspect, the present invention provides a multi-model time-sharing computing power measurement method for an intelligent computing cloud platform, comprising: Step S1: Obtain CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform; Step S2: Calculate the load curve of the object to be measured within the time period using the LSTM model based on the CPU utilization, GPU metric data, storage metric data, and network metric data; output the time period label of the time period to be measured using the random forest classification model based on the actual load data in the GPU metric data, the load curve, and time series features; and output the time period weight and the metric weight of different types of heterogeneous resource data based on the time period label using the XGBoost regression model. Step S3: Based on the time period weight, the indicator weight, and the actual load data, output the first computing power value of the object to be measured using a linear regression model, and convert the first computing power value into a value in units of 1 degree computing power. Step S4: Output energy efficiency factor based on the environmental and energy efficiency index data using the decision tree model; output business factor based on the business index data using the K-Means clustering model; and output load factor based on the load curve and the actual load data. Step S5: Calculate the second computing power value used by the object to be measured during the time period based on the load factor, the business factor, the energy efficiency factor, and the first computing power value.

[0011] In one embodiment, step S4 includes: Step S4.1: Determine the power usage efficiency and load correlation based on the environmental and energy efficiency index data; Step S4.2: Based on the power usage efficiency and the load correlation, the decision tree model outputs the energy efficiency factor.

[0012] In one embodiment, step S4 includes: Step S4.3: Determine the real-time rendering coefficient, batch processing task characteristics, and task priority based on the business indicator data; Step S4.4: Generate the business feature vector based on the real-time rendering coefficients, the batch processing task features, and the task priority; Step S4.5: Output the business factor based on the business feature vector using the K-Means clustering model.

[0013] In one embodiment, step S2 includes: Step S2.1: In response to the deviation between the actual load data and the load curve not being within the optimal load threshold range, or the actual load data being greater than the load threshold, adjust the load threshold and / or the boundary of the time period to be measured.

[0014] In one embodiment, step S2.1 includes: Step S2.1: Based on the actual load data and the load curve, the random forest classification model generates the optimal load threshold range.

[0015] In one embodiment, step S2 includes: Step S2.2: In response to the time period label being a peak time period, the weights of the indicators of GPU floating-point operation count and streaming multiprocessor (SM) utilization are increased through the XGBoost regression model.

[0016] Secondly, the present invention also provides a multi-model time-sharing computing power metering device for an intelligent computing cloud platform, comprising: The acquisition module is used to acquire CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform. The first output module is used to calculate the load curve of the object to be measured in the time period to be measured based on the CPU utilization, GPU index data, storage index data and network index data using an LSTM model; to output the time period label of the time period to be measured based on the actual load data in the GPU index data, the load curve and time series features using a random forest classification model; and to output the time period weight and index weight of different types of heterogeneous resource data based on the time period label using an XGBoost regression model. The second output module is used to output the first computing power value of the object to be measured based on the time period weight, the indicator weight and the actual load data through a linear regression model, and convert the first computing power value into a value in 1 degree computing power. The third output module is used to output energy efficiency factors based on the environmental and energy efficiency index data through a decision tree model, output business factors based on the business index data through a K-Means clustering model, and output load factors based on the load curve and the actual load data. The calculation module is used to calculate the second computing power value used by the object to be measured during the time period based on the load factor, the business factor, the energy efficiency factor, and the first computing power value.

[0017] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps in the multi-model time-sharing computing power metering method of the intelligent computing cloud platform described in the first aspect above.

[0018] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the multi-model time-sharing computing power metering method of the intelligent computing cloud platform described in the first aspect above.

[0019] Fifthly, the present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps in the multi-model time-sharing computing power metering method of the intelligent computing cloud platform as described in the first aspect above.

[0020] In this invention, step S1 involves acquiring CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform; step S2 involves using an LSTM model to calculate the load curve of the object to be measured within the time period based on the CPU utilization, GPU performance data, storage performance data, and network performance data; using a random forest classification model to output the time period label of the time period based on the actual load data in the GPU performance data, the load curve, and time-series features; and using an XGBoost regression model to output the time period weights and different types of heterogeneous data based on the time period label. Step S3: Using a linear regression model, based on the time period weights, the indicator weights, and the actual load data, output the first computing power value of the object to be measured, and convert the first computing power value into a value in units of 1 degree computing power; Step S4: Using a decision tree model, output the energy efficiency factor based on the environmental and energy efficiency indicator data, use a K-Means clustering model, output the business factor based on the business indicator data, and output the load factor based on the load curve and the actual load data; Step S5: Calculate the second computing power value used by the object to be measured during the time period to be measured based on the load factor, the business factor, the energy efficiency factor, and the first computing power value. Thus, addressing the industry pain points of high complexity and lack of unified standards in computing power measurement, this invention integrates CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data from intelligent computing cloud platforms. Based on historical behavior, it performs prediction, real-time load comparison, adaptive allocation of heterogeneous resource weights, time-sharing computing power standardization, and joint energy efficiency-value correction based on load factors, business factors, and energy efficiency factors. This allows for the rapid and accurate removal of interference such as ineffective idle cycles, communication congestion, and inefficient bottlenecks, measuring only the high-efficiency computing that truly generates business value. Furthermore, by using "1 degree of computing power" as a unified unit, it bridges the performance gap between heterogeneous chips such as GPUs, NPUs, and CPUs, making computing power under different architectures, tasks, and energy efficiency levels comparable, tradable, and auditable. This method not only significantly improves the fairness and accuracy of measurement but also embeds green and low-carbon principles and business priority orientation, providing a much-needed and implementable foundation for computing power measurement in integrated intelligent computing cloud platforms. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a multi-model time-sharing computing power measurement method for an intelligent computing cloud platform provided by the present invention; Figure 2 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention; Figure 3 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention; Figure 4 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention; Figure 5 This is a schematic diagram of the overall process of a multi-model time-sharing computing power measurement method for an intelligent computing cloud platform provided by the present invention; Figure 6 This is a schematic diagram of another intelligent computing cloud platform multi-model time-sharing computing power metering system provided by the present invention; Figure 7 This is a structural diagram of a multi-model time-sharing computing power metering device for an intelligent computing cloud platform provided by the present invention; Figure 8 This is a structural diagram of an electronic device provided by the present invention. Detailed Implementation

[0023] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The “computing power” in this invention refers to: the ability of computer equipment or computing / data center to process information; the ability of computer hardware and software to work together to perform a certain computing requirement; the computing power to achieve the target output by processing information data; and a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity, and mainly provides services to society through computing power infrastructure.

[0025] The "Computational Power" (CP) of this invention refers to the ability of a data center server to process data and output results. It is a comprehensive indicator for measuring the computing power of a data center, encompassing general computing power, supercomputing power, and intelligent computing power. The commonly used unit of measurement is floating-point operations per second (FLOPS, 1 EFLOPS = 10^18 FLOPS), with higher values ​​indicating stronger overall computing power. It is estimated that 1 EFLOPS is approximately the computing power output of 5 Tianhe-2A supercomputers, 500,000 mainstream server CPUs, or 2 million mainstream laptops. The calculation formula is: CP = CP 通用 +CP 智能 +CP 超级 .

[0026] The "Network Power" (NP) of this invention refers to the performance of data transmission capability of computing facilities, which includes comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling, and involves network transmission within and between data centers. It is a comprehensive indicator for measuring network transmission scheduling capability.

[0027] The "Storage Power" (SP) of this invention refers to the comprehensive capabilities of a data center in four aspects: data storage capacity, performance, security and reliability, and green and low-carbon operation. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and internal storage devices in servers. The commonly used unit of measurement for storage capacity is exabytes (EB, 1EB = 2^60 bytes), the commonly used unit of measurement for performance is the number of read / write operations per second per unit capacity (IOPS / TB), and the disaster recovery ratio is an important indicator of security and reliability.

[0028] The "computing infrastructure" of this invention refers to a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, enabling centralized computing, storage, transmission, and application of information.

[0029] The "new information infrastructure" of this invention refers to network infrastructure, including 5G networks, fiber optic broadband networks, backbone networks, international communication networks, satellite internet, etc.; computing power infrastructure, including data centers, general computing power centers, intelligent computing centers, supercomputing centers, etc.; and new technology facilities such as artificial intelligence, blockchain, and quantum computing.

[0030] The “computing power” of this invention includes: general computing power, intelligent computing power, and supercomputing power.

[0031] The "general computing power" of this invention refers to the computing power provided by servers based on CPU (Central Processing Unit) chips, which is used to support basic general computing such as cloud computing and edge computing.

[0032] The "intelligent computing power" of this invention refers to: a computing platform deployed on a large scale based on dedicated chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit) for various artificial intelligence innovative applications, such as natural language processing and machine vision.

[0033] The “supercomputing power” of this invention refers to the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for computing in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0034] The "intelligent computing center" of this invention refers to a facility that, through the use of large-scale heterogeneous computing resources, including general-purpose computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.), primarily provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models). The intelligent computing center encompasses facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0035] The "Intelligent Computing Center Cloud Platform" of this invention, abbreviated as "Intelligent Computing Cloud", refers to a cloud computing platform that integrates hardware and software resources based on an intelligent computing center.

[0036] The "intelligent computing center" of this invention includes, but is not limited to, "smart computing center".

[0037] The "intelligent computing center" of this invention, namely the artificial intelligence computing center, is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting an artificial intelligence computing architecture.

[0038] The “computing center” of this invention refers to a facility that is mainly composed of infrastructure such as wind, thermal, hydro, and electricity and IT hardware and software equipment, and has computing power, carrying capacity, and storage capacity, including general data centers, intelligent computing centers, supercomputing centers, etc.

[0039] The "supercomputing center" of this invention refers to a supercomputing data center, which is a data center based on supercomputers or large-scale computing clusters, capable of providing large-scale computing, storage and network services, and widely used in aerospace, defense, oil exploration, climate modeling and genome sequencing and other application scenarios.

[0040] The “computing resources” of this invention refer to the technologies and facilities required for the development of the digital society that have the ability to compute, transmit, store and apply information, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and supporting and guaranteeing resources such as wind, fire, water and electricity.

[0041] The “model” of this invention includes, but is not limited to, “large language model” and “multimodal large model”.

[0042] The "large language model" in this invention refers to a large-scale language model (LLM), which is a language model with a large number of parameters. It is designed to understand and generate human language, and is trained with a large amount of text data. It can perform a wide range of tasks, including text summarization, translation, and sentiment analysis.

[0043] The “Multimodal Large Models” of this invention refer to models that combine multimodal information such as text, images, videos, and audio for training, including but not limited to multimodal large language models.

[0044] The "heterogeneous resources" of this invention refer to a collection of hardware or software resources of different types, architectures, performance characteristics, or functional orientations in an intelligent computing cloud platform. These resources differ significantly in instruction sets, computing models, memory structures, communication methods, and energy efficiency characteristics, making it impossible to schedule, program, or meter them in a unified manner.

[0045] The "degree" in this invention is a unit of measurement for computing power, and its properties and uses are equivalent to "meter", "second", "kilogram", "ampere", "candela", "mole", "Kelvin", etc. The "degree" in this invention is illustrative and not restrictive. Other units can also be used to measure computing power. Those skilled in the art, under the guidance of this invention, can make many extensions without departing from the spirit and scope of the claims, and all such extensions are within the protection scope of this invention.

[0046] The "1-degree computing power" of this invention refers to 1 unit of computing power, which means the ability of the intelligent computing center to meet the model training within a preset time range. The larger the value, the stronger the ability to meet the model training within the preset time range; or, it refers to the computing power provided by the intelligent computing center during model training. The larger the value, the stronger the computing power within the preset time range, or the larger the value, the less time the intelligent computing center needs to complete the same model training effect, that is, the higher the efficiency of model training. Please refer to the applicant's patent documents CN118760575B and CN 119739442 B: The nature and use of the "1 degree of computing power" in this invention are equivalent to "1 degree of electricity". For example, "1 degree of computing power" is 1 unit of computing power, which can be X TFLOPS·h, that is, X trillion floating-point operations per second × hours. "X TFLOPS·h" can be, but is not limited to, any number such as "1 TFLOPS·h", "312 TFLOPS·h", "624 TFLOPS·h", "1024 TFLOPS·h", "10000 TFLOPS·h", etc. The numbers are merely illustrative and not restrictive. Those skilled in the art can make many extensions under the guidance of this invention without departing from the spirit and scope of the claims, and all such extensions are within the protection scope of this invention.

[0047] It is important to emphasize that the difficulty of measuring computing power in this invention is far greater than the difficulty of measuring other fields such as electricity, flow, and storage in existing technologies. Computing power is not a "homogeneous, linearly superimposed" physical quantity, but a "dynamic capability" that is highly dependent on task type, system architecture, software stack, and context. This can be specifically reflected in the following aspects: 1. The measurement of computing power differs fundamentally from the measurement of other fields such as electricity, flow, and storage in existing technologies: computing power is not a standardized commodity. Electricity is energy, flow is information, and storage is space, all of which are scalar quantities. However, computing power is processing capability, which is essentially a multi-dimensional vector (precision, parallelism, memory bandwidth, latency sensitivity, etc.) and is highly nonlinear.

[0048] 2. Modern computing power consists of various chips such as CPU, GPU, TPU, NPU, FPGA, and ASIC. The performance of the same task can vary greatly depending on the hardware used. For example: Image recognition: GPU is 100 times faster than CPU; Encrypted computing: The CPU AES-NI instruction set is much faster than the GPU; Cross-architecture computing power cannot be measured using a single metric (such as FLOPS).

[0049] 3. Computing power measurement is task-dependent; the value of computing power is determined by the task type, and the "effective computing power" of the same hardware varies greatly for different tasks. For example: Scientific computing: Emphasizes double-precision floating-point; AI training: Emphasis on half-precision / mixed precision; Real-time inference: Focus on latency (ms) rather than throughput.

[0050] 4. System coordination overhead is significant in computing power measurement, and actual computing power is severely constrained by non-computational factors. These include memory bandwidth bottlenecks, multi-GPU communication latency, software stack efficiency, and operating system scheduling jitter.

[0051] 5. Computing power measurement is dynamic in time and space. The computing power of existing technologies fluctuates over time, and the computing power of different nodes within the same cluster may be inconsistent (hardware aging, configuration differences, etc.). For example, temperature-induced frequency reduction, power consumption limitations, multi-tenant interference, etc.

[0052] 6. There is a lack of universal benchmarks and units for measuring computing power. Electricity is measured in kWh, electricity in GB, and storage in TB, but computing power lacks a widely accepted "monetary unit." For example, using the same GPU instance for training models and inference tasks should result in different values. When measuring computing power, due to the lack of a unified unit of measurement, it is impossible to describe "how much computing power I used?" The only answer is: "I used 1 hour of A100".

[0053] This would prevent computing power from being "plug-and-play and pay-as-you-go" like electricity, hindering the establishment of computing power grid connection, computing power banks, and computing power trading markets, thus resulting in a lack of unified scheduling basis for cross-domain scheduling.

[0054] Please see Figure 1 , Figure 1 This is a flowchart of a multi-model time-sharing computing power measurement method for an intelligent computing cloud platform provided by the present invention, such as... Figure 1 As shown, the method includes: Step S1: Obtain CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform.

[0055] In intelligent computing cloud platforms, the computing power consumption of objects to be measured (such as tenants, projects, computing tasks, and microservice instances) is not an isolated event, but a dynamic behavior embedded in complex business flows and heterogeneous resource pools. Therefore, it is necessary to reconstruct the true picture of resource usage through multi-dimensional, cross-system, and time-series data collection, so as to provide a data foundation for subsequent accurate computing power measurement, anomaly detection, and cost accounting.

[0056] It should be noted that the CPU utilization rate of the intelligent computing cloud platform reflects the intensity of general computing load and is a basic indicator of computing power consumption.

[0057] It should be noted that CPU utilization can include various data, and no limitation is made here. For example, CPU utilization may include, but is not limited to, the following data: User CPU Utilization: The percentage of CPU used by user-space programs.

[0058] System CPU Utilization: The percentage of CPU used by kernel space programs.

[0059] Idle CPU Utilization: Percentage of CPU idle time.

[0060] I / O Wait CPU Utilization: Percentage of CPU time waiting for I / O to complete.

[0061] Steal CPU Utilization: The percentage of CPU time that a virtual machine "steals" from the host machine (for virtualization environments).

[0062] CPU Context Switches: The number of context switches that occur per second, reflecting system scheduling overhead.

[0063] CPU interrupts: The number of hardware interrupts that occur per second, reflecting the frequency of hardware event processing.

[0064] CPU run queue length: The number of processes / threads waiting for CPU scheduling, reflecting the CPU load.

[0065] CPU Frequency: The current operating frequency of the CPU, which may be affected by dynamic frequency adjustment.

[0066] CPU IPC (Instructions Per Cycle): The number of instructions per cycle, reflecting the CPU's execution efficiency.

[0067] The GPU metrics data of the intelligent computing cloud platform serve as the core carrier of AI computing power, and its "actual load" reflects the real business throughput.

[0068] It should be noted that GPU utilization can include various data, and no limitations are made here. For example, GPU utilization may include, but is not limited to, the following data: GPU Utilization (SM Utilization): Streaming Multiprocessor (SM) utilization, reflecting the busyness of the GPU core computing units.

[0069] Memory Utilization: The percentage of GPU memory that is being used.

[0070] Memory Clock Frequency: The current operating frequency of the GPU memory.

[0071] Graphics Clock Frequency (Core Frequency): The current operating frequency of the GPU core.

[0072] Power Consumption: The current power consumption (watts) of the GPU.

[0073] Temperature: GPU core temperature and memory temperature.

[0074] FP16 / FP32 / FP64 TFlops: Actual floating-point computing power, which can be used as feature input for XGBoost.

[0075] Tensor Core Utilization: Tensor Core utilization (if GPU supports it).

[0076] Encoder / Decoder Utilization: Video codec utilization (if GPU is included).

[0077] PCIe Bandwidth: The bandwidth utilization rate for data transfer between the GPU and the CPU.

[0078] GPU Error Rates: ECC error or other hardware error counts.

[0079] Storage metrics are used to measure the performance, capacity, health status, and efficiency of a storage system, directly impacting data loading speed, task throughput, and overall computing power utilization.

[0080] It should be noted that stored metric data can include various types of data, and no limitations are imposed here. For example, stored metric data may include, but is not limited to, the following: IOPS (Input / Output Operations Per Second): The number of read / write operations per second.

[0081] Throughput: The amount of data read or written per second (MB / s or GB / s).

[0082] Latency: The average latency of read and write operations.

[0083] Disk Utilization: Percentage of disk activity.

[0084] Queue Depth: The number of I / O requests waiting to be processed by disk.

[0085] Free Space: The remaining space on the disk or file system.

[0086] Network metrics data are used to measure the performance, stability, and efficiency of internal and external communication within the cluster.

[0087] It should be noted that stored metric data can include various types of data, and no limitations are imposed here. For example, stored metric data may include, but is not limited to, the following: Bandwidth Utilization: The utilization rate of uplink / downlink network bandwidth.

[0088] Packet Rate: The number of data packets sent / received per second.

[0089] Error Rate / Discard Rate: The ratio of erroneous packets to dropped packets at a network interface.

[0090] Latency: Network transmission delay (e.g., Ping delay).

[0091] Connection Count: The number of active network connections.

[0092] Business metrics data refers to a class of high-level indicators that reflect user task semantics, service quality requirements, usage behavior, and business value.

[0093] It should be noted that business metric data can include various types of data, and no limitations are imposed here. For example, business metric data may include, but is not limited to, the following: Job / Task Type: For example, AI inference, AI training, scientific computing, data cleaning, web services, etc.

[0094] Job Priority: The priority set by the business logic.

[0095] Service Level Objective (SLO) Compliance: Whether the service meets its service level objectives.

[0096] User Count / Request Rate: Number of active users or requests per second (for external services).

[0097] Model Inference Latency: The average latency of AI inference.

[0098] Training Progress / Epochs: The progress of AI training or the number of epochs completed.

[0099] Business Value Score: A score for business value that is assessed manually or automatically.

[0100] Environmental and energy efficiency indicators refer to key performance parameters used to measure the energy consumption, thermal management, carbon emissions, and sustainability of intelligent computing cloud platforms.

[0101] It should be noted that environmental and energy efficiency indicator data may include various types of data, and no limitations are imposed here. For example, environmental and energy efficiency indicator data may include, but is not limited to, the following: PUE (Power Usage Effectiveness): The PUE value of a data center.

[0102] Ambient Temperature / Humidity: The ambient temperature and humidity of the computer room (which may affect equipment heat dissipation and PUE).

[0103] In this embodiment of the invention, a data acquisition agent can be deployed to achieve millisecond-level capture and transmission of massive heterogeneous data using DCGM (for GPUs, 10ms level), perf (for CPUs), SNMP (for networks) and Kafka message queues.

[0104] The beneficial technical effects achieved by step S1 are as follows: The focus of this step is not "data collection," but rather the data and hardware foundation for building a trustworthy computing power economy. It allows the intelligent computing cloud platform to evolve from a "resource lessor" to an "intelligent computing power service provider," making it a platform that provides implementable, measurable, and operable solutions.

[0105] Step S2: Using the LSTM model, based on CPU utilization, GPU metrics, storage metrics, and network metrics, calculate the load curve of the object to be measured within the time period to be measured. Using the random forest classification model, based on the actual load data, load curve, and time series features in the GPU metrics, output the time period label of the time period to be measured. Using the XGBoost regression model, based on the time period label, output the time period weight and the index weights of different types of heterogeneous resource data.

[0106] It should be noted that the core objective of predicting the load curve of the object to be measured during the measurement period is to construct an "expected behavior baseline" to provide data support for subsequent computing power measurement.

[0107] Load curves are semantic predictions of resource consumption patterns during task startup. Their fundamental purpose is not to "guess every CPU percentage," but to establish a model of expected task-level resource behavior, providing a reference for subsequent stage segmentation and supporting the early identification of abnormal behavior.

[0108] By constructing an "expected steady state," a dynamic benchmark is provided for computing power measurement. The "load curve" is the ideal operating baseline for the task in the current context. By comparing it with actual load data, it is possible to determine whether it is efficient, abnormal, or wasteful. This can serve as an anchor point for subsequent dynamic, personalized, and scenario-based measurement.

[0109] Current computing power measurement technologies typically employ uniform measurement standards, such as a "one-size-fits-all" measurement model with artificially set fixed thresholds. In this case, the price is the same during peak and off-peak periods. This approach encourages users to submit tasks immediately, leading to cluster overload (queuing) during the day and a large number of GPUs idling at night (waste).

[0110] In this invention, by fusing dual-load data through a random forest classification model, the output is not only the time period boundary, but also load status labels with scheduling and pricing semantics. These labels can be further mapped to the "computing load level" of the intelligent computing cloud platform, such as: Time period labels: The current or future time period is classified into one of four categories: "peak", "peak", "flat", and "valley".

[0111] Optimal load threshold range: The optimal resource load threshold range recommended by the system corresponding to the current time period label.

[0112] For example, when the GPU utilization rate of the intelligent computing cloud platform is <35% and the green electricity ratio is >70%, the time period is labeled as "computing power valley period".

[0113] In one possible implementation, task-level labels can be aligned with the load level of the intelligent computing cloud platform; for example, user tasks running during "valley time" can enjoy lower-cost computing power.

[0114] In intelligent computing cloud platforms, resources are highly heterogeneous, far exceeding the limitations of a single GPU cluster. Instead, they constitute complex systems composed of multi-dimensional heterogeneous resources, such as heterogeneous computing chips, heterogeneous architectures and instruction sets, and heterogeneous performance characteristics. In the traditional "peak, plateau, and valley" timeframes, the proportion of different types of heterogeneous resources used may vary. For example, during peak periods, the proportion of GPU usage is greater than that of other types of computing power, and also greater than the proportion of GPU usage during other periods.

[0115] The traditional "peak, plateau, valley" system only reflects the intensity of resource use, while the "resource value assessment label" further integrates economic, energy efficiency, and strategic dimensions to form a multi-dimensional value level.

[0116] In this invention, the weights of the indicators corresponding to different time period labels can be different, depending on the actual design needs or the actual situation of the intelligent computing cloud platform. For example, during peak periods, the weights of GPU floating-point computation and SM utilization can be increased and set to 0.55.

[0117] The beneficial technical effects achieved by step S2 are as follows: through multi-model collaboration, the raw resource monitoring data of the intelligent computing cloud platform is transformed into a dynamic computing power weighting system, providing a data foundation for realizing a fair, accurate, green, and tradable new generation of computing power measurement system.

[0118] Step S3: Based on time period weights, indicator weights, and actual load data, output the first computing power value of the object to be measured using a linear regression model, and convert the first computing power value into a value in units of 1 degree computing power.

[0119] In existing technologies, computing power measurement is severely fragmented. Current methods charge based on hardware specifications (e.g., "1 hour of computing power from an A100"), which fails to allow for cross-chip comparisons. Furthermore, measuring by resource utilization multiplied by time may overlook efficiency differences; for example, 80% utilization could represent highly efficient computation or simply idle looping. Charging based on task completion is only suitable for specific scenarios (e.g., large model APIs) and cannot be used for general training / inference tasks.

[0120] In this situation, computing power measurement remains at the level of "physical resource rental" and has not risen to the level of "effective intelligent output", which makes it impossible for it to become a commodity that can be standardized and traded.

[0121] Therefore, this invention proposes a unified unit such as 1 degree of computing power, analogous to the unit of measurement in the power system: 1 kilowatt-hour (kWh) = power × time. More importantly, 1 kWh represents a standardized measure of "work capacity," and its value remains consistent regardless of whether the electricity comes from coal-fired power, hydropower, or wind power.

[0122] Similarly, "1 degree of computing power" should represent "the ability to complete effective intelligent computing" and be comparable across hardware. It should also be interpretable across tasks; whether training models or performing inference tasks, 1 degree of computing power represents a similar "intelligent workload." By establishing a measurement system with 1 degree of computing power as the unified unit, it is possible to achieve computing power accumulation, computing power partitioning, computing power pricing, and other functions, supporting billing by the second, splitting by task stage, and cross-regional aggregation.

[0123] The beneficial technical effect achieved by step S3 is that the computing power value is represented by 1 degree of computing power. "1 degree of computing power" is not only a unit of measurement, but also a unified measure of computing power.

[0124] Step S4: Output energy efficiency factors based on environmental and energy efficiency index data using a decision tree model, output business factors based on business index data using a K-Means clustering model, and output load factors based on load curves and actual load data.

[0125] In this invention, the energy efficiency factor measures whether the energy consumed per unit of computing power is efficient and low-carbon, encouraging users and intelligent computing cloud platforms to use computing resources that are energy efficient, have low PUE, and are rich in green electricity.

[0126] The intelligent computing cloud platform can preset a set of adjustment coefficient rules (which can be dynamically configured by operational strategies) based on the current range, serving as an initial energy efficiency incentive signal. For example, if the PUE drops below 1.2, an energy-saving reward is triggered (the energy efficiency factor adjustment coefficient is adjusted to 0.93); if the PUE rises above 1.5, an energy consumption penalty is triggered (the energy efficiency factor adjustment coefficient is adjusted to 1.12).

[0127] In this invention, the load factor is a dynamic adjustment coefficient in the intelligent computing power metering system, used to measure the actual efficiency of the object being measured in using computing resources within a specific time period. It distinguishes between "efficient utilization" and "inefficient occupancy," ensuring that computing power metering reflects the true computing value, rather than simply time or peak usage.

[0128] In this invention, the availability rate is mapped to a business-meaning incentive / penalty coefficient to guide users in optimizing resource usage. It should be noted that the load factor can have a non-linear response, and efficiency improvements should yield significant returns in critical intervals. It can also be set in conjunction with forecasting.

[0129] In this invention, the business factor is a value adjustment coefficient introduced in intelligent computing power metering, resource scheduling or service pricing systems to reflect the relative importance, priority or commercial value of a computing task in its business scenario.

[0130] In traditional cloud computing, all tasks are typically treated the same: one hour of computing power from an A100 processor equals a fixed price, regardless of whether you're training AI or testing scripts. This leads to problems such as compromised critical business operations, the inability of intelligent computing cloud platforms to achieve differentiated profitability, and a lack of user incentive to optimize. For example, errors in real-time inference for autonomous driving could endanger lives, and a one-second delay in a financial risk control model could result in millions in losses, while a few hours' delay in completing nighttime log analysis has no impact whatsoever.

[0131] Therefore, by introducing business factors, computing power measurement can shift from "billing by resource" to "billing by value".

[0132] In this embodiment of the invention, the K-Means clustering model can be used to process business indicator data to determine business types. Then, business factors are determined based on the business types, thereby mapping the identified business types to quantifiable value adjustment coefficients. For example, autonomous driving perception, surgical robot control, and power grid dispatch have higher business priorities, and their corresponding business factors can be set to 1.15 to 1.25. Developer testing and script verification have lower business priorities, and their corresponding business factors can be set to 0.85 to 0.92.

[0133] It should be noted that the mapping relationship between business types and business factors in the K-Means clustering model can be adjusted according to actual design needs. For example, the business factor value of video conferencing can be temporarily increased during holidays. The mapping relationship between business types and business factors can also be customized by the user; for example, enterprise customers can apply for the "Gold Business" label and be assigned a higher factor. Furthermore, the mapping relationship between business types and business factors can be adjusted according to regional differences; for example, western data centers can provide additional factor rewards for designated "Eastern Data, Western Computing" businesses.

[0134] The beneficial technical effects achieved in step S4 are as follows: By generating energy efficiency factors, business factors, and load factors through the K-Means clustering model, the "effective computing power" can be accurately identified and fairly measured. At the same time, a market-oriented incentive mechanism of "high quality, high price; low quality, low price" is constructed, which can upgrade the original "static, single, and extensive" computing power measurement into a dynamic, multi-dimensional, and value-perceived intelligent evaluation system.

[0135] Step S5: Calculate the second computing power value used by the object to be measured during the time period based on the load factor, business factor, energy efficiency factor and the first computing power value.

[0136] Traditional cloud computing billing typically uses hours or minutes as the smallest unit and assumes that resource usage remains constant throughout the entire period (e.g., "1 GPU × 1 hour"). This solution sets a time period to be metered and a second computing power value within each time period. Since task execution status changes dynamically, calculating sub-computing power values ​​prevents a "one-size-fits-all" metering approach.

[0137] The beneficial technical effects achieved by step S5 are: independently calculating the second computing power value for each time period to be measured, which can support dynamic and non-steady-state tasks, better fits the characteristics of intelligent computing cloud platforms, and provides more accurate computing power measurement results.

[0138] In this invention, step S1 involves acquiring CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform; step S2 involves using an LSTM model to calculate the load curve of the object to be measured within the time period based on the CPU utilization, GPU performance data, storage performance data, and network performance data; using a random forest classification model to output the time period label based on the actual load data, load curve, and time-series features in the GPU performance data; and using an XGBoost regression model to output the time period weight based on the time period label. Step S3: Using a linear regression model based on time period weights, indicator weights, and actual load data, output the first computing power value of the object to be measured, and convert the first computing power value into a value in units of 1 degree computing power; Step S4: Using a decision tree model based on environmental and energy efficiency indicator data, output the energy efficiency factor, using a K-Means clustering model based on business indicator data, output the business factor, and using the load curve and actual load data, output the load factor; Step S5: Calculate the second computing power value used by the object to be measured during the time period based on the load factor, business factor, energy efficiency factor, and the first computing power value. Thus, addressing the industry pain points of high complexity and lack of unified standards in computing power measurement, this invention integrates CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data from intelligent computing cloud platforms. It performs historical behavior prediction, real-time load comparison, adaptive allocation of heterogeneous resource weights, time-sharing computing power standardization, and joint energy efficiency-value correction based on load factors, business factors, and energy efficiency factors. This allows for the rapid and accurate removal of interference such as ineffective idle cycles, communication congestion, and inefficient bottlenecks, measuring only the high-efficiency computing that truly generates business value. Furthermore, by using "1 degree of computing power" as a unified unit, it bridges the performance gap between heterogeneous chips such as GPUs, NPUs, and CPUs, making computing power comparable, tradable, and auditable under different architectures, tasks, and energy efficiency levels. This method not only significantly improves the fairness and accuracy of measurement but also embeds green and low-carbon principles and business priority orientation, providing a much-needed and implementable computing power measurement foundation for integrated intelligent computing cloud platforms.

[0139] Figure 2 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention. In this embodiment of the invention, step S4 includes: Step S4.1: Determine the power usage efficiency and load correlation based on environmental and energy efficiency index data.

[0140] Step S4.2: Output the energy efficiency factor based on power usage efficiency and load correlation using a decision tree model.

[0141] It should be noted that energy efficiency is an indicator that measures the effective computational output generated per unit of energy consumption. In intelligent computing and AI task scenarios, it is used to evaluate "how much useful intelligent computing work can be done with one kilowatt-hour of electricity".

[0142] Load correlation refers to the relationship between the current system load level and PUE. System load level refers to the actual computing / operational intensity of intelligent computing cloud platform devices (such as servers, GPUs, storage, etc.).

[0143] Energy efficiency has become a core dimension for measuring computing power quality, and the traditional measurement method of "only looking at computing power output and not energy consumption cost" is no longer sustainable.

[0144] High-efficiency computing power is not equivalent to high-energy-consuming computing power; energy efficiency must be incorporated into the computing power value assessment system. This invention can construct a "green computing power value" through a coupling mechanism of power usage efficiency and load correlation.

[0145] In this embodiment of the invention, step S4.1 involves determining power usage efficiency and load correlation based on environmental and energy efficiency index data; step S4.2 involves outputting an energy efficiency factor based on power usage efficiency and load correlation using a decision tree model. In this way, by integrating power usage efficiency (PUE) and load correlation, a dynamic, interpretable, and guideable energy efficiency factor is generated, enabling computing power metering to move from "extensive energy consumption allocation" to "refined green pricing," thereby improving data center energy efficiency and guiding user behavior towards a low-carbon and high-efficiency direction.

[0146] Figure 3 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention. In this embodiment of the invention, step S4 includes: Step S4.3: Determine the real-time rendering coefficient, batch processing task characteristics, and task priority based on business indicator data.

[0147] Step S4.4: Generate business feature vectors based on real-time rendering coefficients, batch processing task features, and task priorities.

[0148] Step S4.5: Output business factors based on business feature vectors using the K-Means clustering model.

[0149] In this invention, the real-time rendering coefficient is a value that measures the sensitivity of a service to latency (such as end-to-end latency, frame rate stability, and interaction response speed).

[0150] Batch processing task characteristics are used to describe whether a task is a large-scale, non-interactive, interruptible computational job without a strict deadline, reflecting its low dependence on "real-time".

[0151] Task priority describes the relative importance of a task in platform scheduling and resource allocation, and is usually determined by SLA, tenant level, and scope of business impact.

[0152] After obtaining the above three types of information, the above three types of features can be normalized, encoded and concatenated into a multi-dimensional vector.

[0153] Figure 4 This is a flowchart of another multi-model time-sharing computing power measurement method for intelligent computing cloud platforms provided by the present invention. In this embodiment of the present invention, step S2 includes: Step S2.1: In response to the deviation between the actual load data and the load curve not being within the optimal load threshold range, or the actual load data being greater than the load threshold, adjust the load threshold and / or the boundary of the time period to be measured.

[0154] In this invention, a machine learning model (random forest) can be used to automatically learn "typical load patterns" from historical and real-time load data, and generate adaptive load thresholds accordingly. When new data deviates significantly from expectations, online adjustments to the thresholds or time period boundaries are triggered, enabling continuous tracking and response to changes in task behavior.

[0155] In this embodiment of the invention, the optimal load threshold range is the optimal resource load threshold range recommended by the system corresponding to the current time period label.

[0156] When the deviation value is outside the optimal load threshold range, or when the actual load data exceeds the load threshold, the system dynamically adjusts the load threshold and / or the boundary of the time period to be measured, and feeds this adjustment back to the model and time slicing mechanism, thus forming a closed-loop adaptive control system of "perception, judgment, adjustment, verification, and relearning". This closed loop is not only a technical implementation detail, but also the key to the continuous evolution capability of the entire intelligent computing power metering system.

[0157] Because the resource usage of intelligent computing cloud platforms is highly non-stationary, GPU utilization may fluctuate significantly, resulting in severely distorted average utilization. Current technologies use static billing based on "GPU hours," causing users to pay for "waiting time." Furthermore, computing power output depends not only on GPUs but also on factors such as storage IOPS, network throughput, and CPU data preprocessing. Bottlenecks in any of these areas can cause GPUs to "idle," but traditional metering methods cannot identify this.

[0158] In this invention, by constructing a load curve, the time-series trajectory of the task's true "effective computational intensity" can be reflected. Simultaneously, by determining the "optimal load threshold range," exceeding this range can be considered either a waste of resources or inefficiency. In one possible implementation, real-time comparison and dynamic adjustment can be performed as follows: If the deviation is too large (e.g., the load curve is flat but there is a sudden spike), then the time period boundary is extended to capture the complete behavior.

[0159] If the threshold is consistently exceeded (e.g., >95% for an extended period), the measurement granularity will be reduced or sub-window splitting will be triggered to avoid overload distortion.

[0160] If the temperature remains below the threshold for an extended period, adjacent inefficient windows can be merged to reduce noise interference.

[0161] In this invention, the optimal load threshold range can be pre-set, for example, manually configured by maintenance personnel based on historical data, and can be changed according to actual design needs.

[0162] The optimal load threshold range can also be dynamically adjusted to ensure that each computing power measurement is aligned with the actual value output cycle of the task, laying a key technological foundation for building a fair, efficient, green, and tradable national integrated computing power network.

[0163] In another possible implementation, step S2.1 further includes the following steps: Step S2.1: Generate the optimal load threshold range based on the actual load data and load curve using the random forest classification model.

[0164] In the invention, step S2 further includes the following steps: Step S2.2: In response to the time period label being a peak time period, the weights of the indicators of GPU floating-point operation count and streaming multiprocessor (SM) utilization are increased through the XGBoost regression model.

[0165] During peak periods, users often run tasks with high SLA and high revenue relevance. These tasks are highly dependent on the computing power of GPUs. By increasing the weight of metrics such as GPU floating-point operations and streaming multiprocessor (SM) utilization, computing power measurement can focus on the parts that truly create business value, thus alleviating computing power pressure during peak periods.

[0166] Figure 5 This is a schematic diagram of the overall process of a multi-model time-sharing computing power measurement method for an intelligent computing cloud platform provided by the present invention, as shown below. Figure 5 As shown, it includes the following steps: Step S501: Obtain CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform.

[0167] Step S502: Calculate the load curve of the object to be measured during the time period using the LSTM model based on CPU utilization, GPU metrics, storage metrics, and network metrics.

[0168] Step S503: Based on the actual load data, load curve and time series characteristics in the GPU index data, output the time period label of the time period to be measured using the random forest classification model.

[0169] Step S504: Using the XGBoost regression model based on time period labels, output the time period weights and indicator weights for different types of heterogeneous resource data.

[0170] Step S505: Detect the deviation between the actual load data and the load curve.

[0171] Step S506: In the event of abnormal deviation, adjust the load threshold and / or the boundary of the time period to be measured, and return to re-execute step S502 and its subsequent steps.

[0172] Step S507: Under normal deviation conditions, the first computing power value of the object to be measured is output through the linear regression model based on the time period weight, indicator weight and actual load data, and the first computing power value is converted into a value in units of 1 degree computing power.

[0173] Step S508: Output energy efficiency factors based on environmental and energy efficiency index data using a decision tree model, output business factors based on business index data using a K-Means clustering model, and output load factors based on load curves and actual load data.

[0174] Step S509: Calculate the second computing power value used by the object to be measured during the time period based on the load factor, business factor, energy efficiency factor and the first computing power value.

[0175] Figure 6 This is a schematic diagram of another intelligent computing cloud platform multi-model time-sharing computing power metering system provided by the present invention, as shown below. Figure 6 As shown, the multi-model time-sharing computing power metering system of this intelligent computing cloud platform includes a data acquisition layer, an algorithm layer, a processing layer, a business layer, and a display layer.

[0176] It should be noted that the acquisition layer includes the function of unified access to multiple heterogeneous sources. It decouples production and consumption through Kafka queues, thereby supporting high-concurrency and low-latency data stream processing.

[0177] The algorithm layer is responsible for extracting semantic, pattern, and prediction information from the raw data. It achieves end-to-end automation, enabling business recognition without manual annotation.

[0178] The processing layer transforms the "model output" into an "executable factor." When the random forest detects anomalous behavior, it triggers a "dynamic threshold" adjustment. The new threshold is fed back to K-Means and LSTM, forming an adaptive learning loop.

[0179] The business layer is used to transform technical outputs into actionable business value. The business layer supports, but is not limited to, the following functionalities: Support "pay-per-value": high-value tasks enjoy premiums or guarantees.

[0180] Support "green incentives": tasks with high energy efficiency factors receive discounts.

[0181] Supports "Resource Optimization Suggestions": Prompts users to optimize resources based on load factors.

[0182] The presentation layer provides transparent and intuitive visualization services to users and operations and maintenance personnel.

[0183] The multi-model time-sharing computing power metering system of the intelligent computing cloud platform is not only a "computing power billing system," but also an intelligent operating system for AI asset operation. This system can realize the process from data flow collection to processing to decision-making, the process from value flow resources to efficiency to price, and the process from control flow perception to adjustment to optimization. Through the three pillars of machine learning, dynamic factors, and closed-loop feedback, it achieves a paradigm shift from "resource leasing" to "intelligent services," providing a solid technical foundation for building a green, fair, efficient, and reliable nationwide integrated computing power network.

[0184] Please see Figure 7 , Figure 7 This is a structural diagram of a multi-model time-sharing computing power metering device for an intelligent computing cloud platform provided by the present invention, as shown in the figure. Figure 7 As shown, the multi-model time-sharing computing power metering device 700 of the intelligent computing cloud platform includes: The acquisition module 701 is used to acquire CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform.

[0185] The first output module 702 is used to calculate the load curve of the object to be measured within the time period based on CPU utilization, GPU indicator data, storage indicator data and network indicator data using an LSTM model; to output the time period label of the time period based on the actual load data, load curve and time series features in the GPU indicator data using a random forest classification model; and to output the time period weight and indicator weight of different types of heterogeneous resource data based on the time period label using an XGBoost regression model.

[0186] The second output module 703 is used to output the first computing power value of the object to be measured based on the time period weight, indicator weight and actual load data through a linear regression model, and convert the first computing power value into a value in units of 1 degree computing power.

[0187] The third output module 704 is used to output energy efficiency factors based on environmental and energy efficiency index data through a decision tree model, output business factors based on business index data through a K-Means clustering model, and output load factors based on load curves and actual load data.

[0188] The calculation module 705 is used to calculate the second computing power value used by the object to be measured during the time period based on the load factor, business factor, energy efficiency factor and the first computing power value.

[0189] The third output module 704 includes: The first determining unit is used to determine the power usage efficiency and load correlation based on environmental and energy efficiency index data.

[0190] The first output unit is used to output the energy efficiency factor based on the power usage efficiency and load correlation through the decision tree model.

[0191] The third output module 704 includes: The second determining unit is used to determine the real-time rendering coefficients, batch processing task characteristics, and task priorities based on business indicator data.

[0192] The generation unit is used to generate business feature vectors based on real-time rendering coefficients, batch processing task characteristics, and task priorities. The second output unit is used to output business factors based on business feature vectors using the K-Means clustering model.

[0193] The first output module 702 includes: The adjustment unit is used to adjust the load threshold and / or the boundary of the time period to be measured in response to the deviation between the actual load data and the load curve being outside the optimal load threshold range, or the actual load data being greater than the load threshold.

[0194] The adjustment unit is also used for: The optimal load threshold range is generated by using a random forest classification model based on actual load data and load curves.

[0195] The first output module 702 is also used for: In response to the peak time period label, the weights of the metrics for GPU floating-point operations and streaming multiprocessor (SM) utilization are increased through the XGBoost regression model.

[0196] The multi-model time-sharing computing power metering device for the intelligent computing cloud platform provided by this invention can realize the various processes of the various embodiments of the above-mentioned multi-model time-sharing computing power metering method for the intelligent computing cloud platform. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0197] It should be noted that the multi-model time-sharing computing power metering device of the intelligent computing cloud platform in this invention can be a device, or it can be a component, integrated circuit, or chip in an electronic device.

[0198] The present invention also provides an electronic device, see below. Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device includes a memory 801, a processor 802, and a program or instructions stored in the memory 801 that run on the processor 802. When the program or instructions are executed by the processor 802, they can achieve the following: Figure 1 The steps in the corresponding multi-model time-sharing computing power metering method embodiment of the intelligent computing cloud platform and the achievement of the same beneficial effect will not be repeated here.

[0199] The processor 802 can be a CPU, ASIC, FPGA, or GPU.

[0200] Those skilled in the art will understand that all or part of the steps of the above-described embodiments of the multi-model time-sharing computing power metering method for intelligent computing cloud platforms can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.

[0201] The present invention also provides a readable storage medium on which a computer program is stored, and which, when executed by a processor, can perform the above-described functions. Figure 1 Any step in the corresponding multi-model time-sharing computing power metering method embodiment of the intelligent computing cloud platform can achieve the same technical effect, and will not be described again here to avoid repetition. Storage media, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, etc.

[0202] The present invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the above-described... Figure 1 The various processes of the corresponding multi-model time-sharing computing power metering method implementation for intelligent computing cloud platforms can achieve the same technical effect, and will not be described again here to avoid repetition.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or second terminal device, etc.) to execute the methods of the various embodiments of the present invention.

[0204] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A multi-model time-sharing computing power measurement method for an intelligent computing cloud platform, characterized in that, include: Step S1: Obtain CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform; Step S2: Calculate the load curve of the object to be measured within the time period using the LSTM model based on the CPU utilization, GPU metric data, storage metric data, and network metric data; output the time period label of the time period to be measured using the random forest classification model based on the actual load data in the GPU metric data, the load curve, and time series features; and output the time period weight and the metric weight of different types of heterogeneous resource data based on the time period label using the XGBoost regression model. Step S3: Based on the time period weight, the indicator weight, and the actual load data, output the first computing power value of the object to be measured using a linear regression model, and convert the first computing power value into a value in units of 1 degree computing power. Step S4: Output energy efficiency factor based on the environmental and energy efficiency index data using the decision tree model; output business factor based on the business index data using the K-Means clustering model; and output load factor based on the load curve and the actual load data. Step S5: Calculate the second computing power value used by the object to be measured during the time period based on the load factor, the business factor, the energy efficiency factor, and the first computing power value.

2. The method according to claim 1, characterized in that, Step S4 includes: Step S4.1: Determine the power usage efficiency and load correlation based on the environmental and energy efficiency index data; Step S4.2: Based on the power usage efficiency and the load correlation, the decision tree model outputs the energy efficiency factor.

3. The method according to claim 2, characterized in that, Step S4 includes: Step S4.3: Determine the real-time rendering coefficient, batch processing task characteristics, and task priority based on the business indicator data; Step S4.4: Generate the business feature vector based on the real-time rendering coefficients, the batch processing task features, and the task priority; Step S4.5: Output the business factor based on the business feature vector using the K-Means clustering model.

4. The method according to any one of claims 1-3, characterized in that, Step S2 includes: Step S2.1: In response to the deviation between the actual load data and the load curve not being within the optimal load threshold range, or the actual load data being greater than the load threshold, adjust the load threshold and / or the boundary of the time period to be measured.

5. The method according to claim 4, characterized in that, Step S2.1 includes: Step S2.1: Based on the actual load data and the load curve, the random forest classification model generates the optimal load threshold range.

6. The method according to claim 1, characterized in that, Step S2 includes: Step S2.2: In response to the time period label being a peak time period, the weights of the indicators of GPU floating-point operation count and streaming multiprocessor (SM) utilization are increased through the XGBoost regression model.

7. A multi-model time-sharing computing power metering device for an intelligent computing cloud platform, characterized in that, include: The acquisition module is used to acquire CPU utilization, GPU performance data, storage performance data, network performance data, business performance data, and environmental and energy efficiency performance data of the intelligent computing cloud platform. The first output module is used to calculate the load curve of the object to be measured in the time period to be measured based on the CPU utilization, GPU index data, storage index data and network index data using an LSTM model; to output the time period label of the time period to be measured based on the actual load data in the GPU index data, the load curve and time series features using a random forest classification model; and to output the time period weight and index weight of different types of heterogeneous resource data based on the time period label using an XGBoost regression model. The second output module is used to output the first computing power value of the object to be measured based on the time period weight, the indicator weight and the actual load data through a linear regression model, and convert the first computing power value into a value in 1 degree computing power. The third output module is used to output energy efficiency factors based on the environmental and energy efficiency index data through a decision tree model, output business factors based on the business index data through a K-Means clustering model, and output load factors based on the load curve and the actual load data. The calculation module is used to calculate the second computing power value used by the object to be measured during the time period based on the load factor, the business factor, the energy efficiency factor, and the first computing power value.

8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the multi-model time-sharing computing power metering method for the intelligent computing cloud platform as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A computer program is stored on a computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the multi-model time-sharing computing power metering method for the intelligent computing cloud platform as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of the multi-model time-sharing computing power metering method for the intelligent computing cloud platform as described in any one of claims 1 to 6.

Images