Artificial intelligence platform resource metering management method and system, medium and processor

Abstract

The invention provides an artificial intelligence platform resource metering management method which specifically comprises the following steps: S1, detecting and metering the computing power surplus of each position computing power point in a platform system, and dividing the computing power points into training points and reasoning points according to the idle computing power surplus and computing power point types; s2, detecting and metering the storage performance index of each position storage point in the platform system to obtain the parameter of each storage point; s3, evaluating an algorithm model in the platform system; s4, detecting and evaluating the service data in the platform system; and S5, selecting a computing power point, a storage point, an algorithm model and business data to form a task combination, so that the utilization rate and benefit of various resources are optimal. According to the scheme, the computing power and the network resources are integrated into a whole, an efficient, extensible and flexible ubiquitous computing power network is formed, the cluster advantage of the computing power is exerted, the scale efficiency of the computing power is improved, and global intelligent scheduling and optimization are carried out on the computing network resources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence platform optimization technology, and in particular to an artificial intelligence platform resource metering management method, system, medium and processor. Background Art

[0002] In the technological wave of the 21st century, artificial intelligence (AI), as a key force leading the development of future science and technology, is changing our lives, work and even the entire society at an unprecedented speed. The vigorous development of artificial intelligence is inseparable from the support of three core engines: algorithms, computing power and computing materials (i.e. data). These three complement each other and jointly promote the continuous breakthroughs and innovations in artificial intelligence technology. This article will deeply explore the role, current status and future trends of these three engines in the development of artificial intelligence.

[0003] The relationship between computing power and algorithms is mutually reinforcing and mutually restrictive. Computing power provides the basis for algorithm execution, while excellent algorithms can use computing power resources more efficiently. Algorithms are the soul that supports the entire intelligent world, and computing power is the physical basis for carrying algorithms.

[0004] How to uniformly manage and schedule computing resources with different characteristics, balance the load of each node, and dynamically allocate resources to achieve more fine-grained resource management, improve computing resource utilization, and achieve the goal of reducing costs and increasing efficiency. With the new trend of the development of artificial intelligence platforms, there are still few studies on the metering and optimization of platform resources, and the platform's resource utilization is low. Most of the existing research is still in the small-scale optimization within a single system.

[0005] In view of this, a method, system, medium and processor for resource metering management of an artificial intelligence platform are needed. Summary of the invention

[0006] In view of the fact that there is still little research on the metering optimization of platform resources in the prior art and the problem of low resource utilization of the platform, the present invention provides an artificial intelligence platform resource metering management method, system, medium and processor, which can achieve the optimal utilization and benefit of IT resources (computing resources, storage resources, etc.) and atomic capabilities (data sets, model libraries, etc.). The specific technical solution is as follows:

[0007] A method for resource metering management of an artificial intelligence platform, comprising the following steps:

[0008] S1: Detect and measure the surplus computing power of computing power points at various locations in the platform system, and divide the computing power points into training points and inference points according to the surplus idle computing power and computing power point types;

[0009] S2: Detect and measure the storage performance indicators of each storage point in the platform system to obtain the parameters of each storage point;

[0010] S3: Evaluate the algorithm model within the platform system and determine the type, performance indicators and status of the algorithm model;

[0011] S4: Detect and evaluate the business data type, nature and purpose, location distance or data volume within the platform system;

[0012] S5: According to the needs of the business scenario and the evaluation status of each resource, select computing power points, storage points, algorithm models and business data to form a task combination to optimize the utilization and benefits of various resources.

[0013] Furthermore, the computing power points include the same type of computing power points and heterogeneous computing power points; the computing power surplus amount detection and measurement formula of the same type of computing power points is as follows:

[0014] A = a*b*(1-c);

[0015] In the above formula, A is the surplus computing power of a certain type of processor; a is the performance of a single processor; b is the number of processors; and c is the utilization rate of each processor.

[0016] Furthermore, the calculation formula for detecting the surplus amount of computing power of the heterogeneous computing power point is as follows:

[0017]

[0018] In the above formula, A i is the computing power surplus of the i-th type of processor; Y is the computing power surplus of the heterogeneous computing power point; n is the number of processor types in the heterogeneous computing power point.

[0019] Furthermore, the storage performance indicators of the storage point include IOPS, throughput, latency, bandwidth, data protection and storage margin.

[0020] Furthermore, the process of selecting computing power points, storage points, algorithm models and business data to form a task combination according to the needs of the business scenario and the evaluation status of each resource includes the following steps:

[0021] S51: Decompose the tasks according to the needs of the business scenario and clarify the specific task requirements;

[0022] S52: Select a suitable algorithm model according to the decomposed tasks and the characteristics of the data to be used;

[0023] S53: Select a computing power point according to the state of the algorithm model;

[0024] S54: Select storage points and business data based on the available storage points and business data according to the computing power point, taking into account bandwidth, distance and data throughput factors.

[0025] Furthermore, step S54 specifically includes the following steps:

[0026] First, determine the business data that can be used based on the nature of the task;

[0027] Secondly, determine the storage points that can be used based on the scale of data calculation;

[0028] Finally, select specific storage points and business data based on distance, bandwidth speed, and quantity.

[0029] Furthermore, the calculation formula for selecting specific storage points and business data is as follows:

[0030]

[0031] In the above formula, X is the selection ratio; L is the distance between the storage point or business data and the computing power point; L P is the average distance between available storage points or business data and computing power points; D is the bandwidth between storage points or business data and computing power points; D P is the average bandwidth between available storage points or business data and computing power points; T is the data throughput of storage points or business data; T P It is the average throughput of storage points or business data.

[0032] Furthermore, when selecting storage points, in order to consider the impact of storage margin on business expansion, that is, after the current business is completed, when the same business is started next time, there is a high probability that the stored data generated this time will be used, and new and more data will be generated at the same time, or as the calculation progresses, there is a certain probability that new data will be generated according to the progress. At this time, considering data access at the same storage point can improve the reliability of network data use. The specific calculation formula is as follows:

[0033]

[0034] In the above formula, X c Select the weight for the storage point; L is the distance between the storage point or business data and the computing power point; L P is the average distance between available storage points or business data and computing power points; D is the bandwidth between storage points or business data and computing power points; D P is the average bandwidth between available storage points or business data and computing power points; T is the data throughput of storage points or business data; T P is the average throughput of the storage point or business data; C is the storage space margin of the storage point; S is the amount of business data used in the calculation.

[0035] An artificial intelligence platform resource metering management system, applied to the artificial intelligence platform resource metering management method described above, comprises:

[0036] The computing power point measurement module is used to detect and measure the computing power surplus of computing power points at various locations in the platform system, and divide the computing power points into training points and inference points according to the idle computing power surplus and computing power point type;

[0037] The storage point metering module is used to detect and measure the storage performance indicators of storage points at various locations in the platform system to obtain the parameters of each storage point;

[0038] Model measurement module, which is used to evaluate the algorithm model in the platform system and determine the type, performance indicators and status of the algorithm model;

[0039] Business data measurement module, which is used to detect and evaluate the business data type, nature and purpose, location distance or data volume within the platform system;

[0040] The task combination module is used to select computing power points, storage points, algorithm models and business data to form a task combination based on the needs of the business scenario and the evaluation status of each resource, so as to optimize the utilization and benefits of various resources.

[0041] A computer-readable storage medium, the computer-readable storage medium comprising a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the artificial intelligence platform resource metering management method described above.

[0042] A processor is used to run a program, wherein the program executes the artificial intelligence platform resource metering management method described above when running.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. In this application, computing power and network resources are integrated into a whole to form an efficient, scalable and flexible ubiquitous computing power network, so that the platform can evolve towards the integration of ubiquitous interconnection and ubiquitous computing unified scheduling and unified control. By driving the network to perceive the location of computing power, an integrated platform service of local diversion and computing-network integration is realized.

[0045] 2. Ubiquitous computing power connected through the network can break through the performance limit of single-point computing power, give full play to the cluster advantages of computing power, and improve the scale efficiency of computing power. Through global intelligent scheduling and optimization of computing network resources, it can effectively promote the "flow" of computing power and meet the business needs for on-demand use of computing power. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0047] Figure 1 It is a flowchart of a resource metering management method for an artificial intelligence platform;

[0048] Figure 2 The figure is a structural diagram of a resource metering management system for an artificial intelligence platform. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0051] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0052] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0053] Improving the utilization rate of computing resources has a huge impact on the scale and cost of computing power. A McKinsey research report shows that the average daily utilization rate of global servers is usually only 6% at most; according to Gartner statistics, the utilization rate of global data centers is less than 12%. The above data all show that there is a huge "waste" in the server costs and resource consumption of data centers. If the comprehensive utilization rate of computing resources can be increased from 6% to 90%, it means that the macro computing power can be increased by 15 times immediately, and the unit computing power cost will be reduced to 1 / 15.

[0054] Embodiment 1

[0055] like Figure 1 The figure shows a flow chart of a resource metering management method for an artificial intelligence platform, which specifically includes the following steps:

[0056] S1: Detect and measure the surplus computing power of computing power points at various locations in the platform system, divide the computing power points into training points and inference points according to the surplus idle computing power and computing power point types, and form computing power resource pools respectively, and obtain the location distance of each computing power point.

[0057] Furthermore, the types of computing power points include CPU, GPU, FPGA, DSA, NPU or DPU; for example, GPU is more suitable for graphics visualization, rendering, AI reasoning and AI training calculations, FPGA is more suitable for chip design testing, gene sequencing and transcoding calculations, NPU is more suitable for machine learning, etc. At the same time, the types of computing power points also include, for example, the combination of FPGA and CPU and other computing architectures to form CPU+FPGA heterogeneous computing power points, and CPU+GPU+DPU composition to provide rendering, reasoning and computing integrated heterogeneous computing power points, so that the most suitable architecture can do its best computing and acceleration to achieve system optimization.

[0058] The performance of many individual chips is aggregated into a large computing resource pool. Conversely, if the performance of each independent chip cannot be aggregated into a huge computing resource pool, forming isolated islands, no matter how high the performance of a single chip is, it will be meaningless, and it will be difficult to improve the utilization rate.

[0059] To connect individual resources into a huge resource pool, you need to:

[0060] The hardware itself needs to support (hardware) virtualization, such as Intel's VT-x / VT-d technology. IO devices, including IO devices, need to support full hardware virtualization based on technologies such as SR-IOV. The accelerator itself also needs to support virtualized logical processing channels.

[0061] On top of that, virtualization technology can be used to improve the utilization of computing resources such as a single processing chip, and the software migration function in virtualization technology can make it easy for upper-level business software to select different physical resources (in the entire resource pool) to run. This is how resources can be segmented into single hardware resources and pooled into multiple hardware resources.

[0062] The calculation formula for detecting the surplus computing power of the same type of computing power points is as follows:

[0063] A = a*b*(1-c);

[0064] In the above formula, A is the surplus computing power of a certain type of processor; a is the performance of a single processor; b is the number of processors; and c is the utilization rate of each processor.

[0065] The calculation formula for detecting the surplus computing power of the heterogeneous computing power point is as follows:

[0066]

[0067] In the above formula, A i is the computing power surplus of the i-th type of processor; Y is the computing power surplus of the heterogeneous computing power point; n is the number of processor types in the heterogeneous computing power point.

[0068] S2: Detect and measure the storage performance indicators of each storage point in the platform system to obtain the parameters of each storage point, including location distance, storage margin, latency, bandwidth, throughput, IOPS and other indicators.

[0069] The main indicators of storage resource performance include:

[0070] IOPS (Input / Output Operations Per Second): IOPS is an indicator to measure the random read / write performance of a storage system, indicating the number of input / output operations that a storage system can process per second. The higher the IOPS, the better the random read / write performance of the storage system.

[0071] Throughput: Throughput is an indicator to measure the sequential read and write performance of a storage system, indicating the amount of data that a storage system can transmit per second. The higher the throughput, the better the sequential read and write performance of the storage system.

[0072] Latency: Latency is a measure of the storage system's read and write response time, indicating the time it takes for the storage system to process a read or write request. The lower the latency, the shorter the storage system's read and write response time.

[0073] Bandwidth: Bandwidth is an indicator that measures the data transmission speed of a storage system. The higher the bandwidth, the faster the data transmission speed of the storage system.

[0074] Data protection: Data protection is an indicator to measure the data protection capability of a storage system. The use of technologies such as RAID can improve the data protection capability of a storage system, thereby improving the stability and reliability of storage performance.

[0075] Storage margin: Storage margin is an indicator to measure the remaining storage space of the storage system. The larger the storage margin, the lower the utilization of the storage system and the more available space.

[0076] Furthermore, the detection and measurement of the storage performance index may adopt the following evaluation method:

[0077] Use professional tools for testing: You can use professional performance testing tools such as Iometer, IOzone, and FIO. These tools can help measure performance indicators such as IOPS, throughput, and latency.

[0078] Simulate actual application scenarios: Test in actual application scenarios to more accurately evaluate the performance of the storage system. For example, the actual performance of the storage system can be tested by simulating actual application scenarios such as database operations and file system access.

[0079] Consider system configuration and external environmental factors: When testing, you need to consider the impact of system configuration and external environmental factors on performance. For example, factors such as the ratio of reads and writes, access queue depth, and data block size will affect the IOPS results.

[0080] Through the above methods and indicators, the performance of storage resources can be comprehensively evaluated to ensure that they meet the needs of actual applications.

[0081] S3: Evaluate the algorithm model within the platform system to determine the type, performance indicators and status of the algorithm model.

[0082] Furthermore, the state includes, for example, whether it is a model to be trained or an inference model that has been trained and can be used.

[0083] Furthermore, the model library refers to configuring different models to form a whole, and effectively managing and using each model through the model library management system. The models in the model library are sealed and ready for use. Models with different performance can be called for training and use according to different user needs. For example, algorithm models include the following types:

[0084] Linear regression model is used to predict the value of continuous variables. It predicts the value of the target variable by fitting a straight line or hyperplane. Linear regression model is simple and easy to understand and suitable for most data sets, but it has weak modeling ability for nonlinear relationships and is sensitive to outliers and noise.

[0085] Logistic regression model is used to solve binary classification problems. Logistic regression predicts the category of a sample by mapping the output of linear regression to a probability value between 0 and 1. It is suitable for binary classification problems, such as spam classification, disease diagnosis, etc. The output result is a probability, which can intuitively explain the possibility that the sample belongs to the positive class.

[0086] Decision tree model, used for classification and regression tasks. Decision trees divide data sets into different categories or prediction values ​​through a series of feature partitioning. Decision trees are easy to understand and explain, suitable for handling nonlinear relationships and missing values, but prone to overfitting and underfitting, and sensitive to outliers and noise.

[0087] Random forest model, an ensemble learning algorithm based on decision trees. It builds multiple decision trees and makes predictions by voting or averaging. Random forest improves the accuracy and robustness of the model and is suitable for image classification, credit scoring and other fields. It has good generalization ability and anti-overfitting ability.

[0088] Support vector machine model, used for classification and regression tasks. Support vector machine maps data into a high-dimensional space and finds an optimal hyperplane to maximize the interval between different categories. Support vector machine is suitable for processing nonlinear relationships and high-dimensional data, with high accuracy and robustness, but has poor processing capabilities for large-scale data sets.

[0089] Naive Bayes model, a classification algorithm based on Bayes' theorem. Naive Bayes assumes that features are independent of each other and performs classification by calculating posterior probabilities. It is suitable for processing missing values ​​and noisy data, but performs poorly when there is a strong correlation between features.

[0090] Furthermore, the performance of the model is mainly measured by the following indicators:

[0091] Accuracy: Accuracy is a basic indicator for measuring model performance. It intuitively reflects the accuracy of the model by calculating the ratio of the number of instances correctly predicted by the model to the total number of instances in the data set. However, when dealing with unbalanced data sets, the accuracy may appear inaccurate due to the large number of categories.

[0092] Precision: Precision focuses on the accuracy of the model's predictions for positive samples, and calculates the proportion of instances that the model predicts to be positive samples that are actually positive samples. In applications such as medical diagnosis or fraud detection, precision is particularly important because it can reduce the serious consequences of false positives.

[0093] Recall: Recall measures the model's ability to identify positive samples, that is, the proportion of instances that are actually positive samples that the model correctly predicts as positive samples. In scenarios where the model recall rate is of concern, recall is an important indicator.

[0094] F1 Score: F1 score is the harmonic mean of precision and recall, which takes both into account and is suitable for scenarios where both are important. A high F1 score indicates that the model has achieved a good balance between precision and recall.

[0095] Confusion Matrix: A confusion matrix is ​​a two-dimensional table that shows the relationship between the actual category and the model prediction category, including true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The confusion matrix can be used to calculate indicators such as precision and recall.

[0096] ROC curve (Receiver Operating Characteristic Curve): The ROC curve uses the false positive rate (FPR) as the horizontal axis and the true positive rate (TPR) as the vertical axis to plot the performance under different thresholds. The larger the AUC (area under the ROC curve), the better the performance of the model.

[0097] Mean Squared Error (MSE) and Root Mean Squared Error (RMSE): In regression problems, MSE and RMSE are commonly used evaluation indicators that measure the difference between the model's predicted values ​​and the actual values.

[0098] R Squared(R 2 ): R 2 For regression problems, it indicates the ability of the model to explain variation. 2 The closer it is to 1, the stronger the explanatory power of the model.

[0099] S4: Detect and evaluate the business data type, nature and purpose, location distance, data volume and quality within the platform system.

[0100] The evaluation of data quality includes the following main indicators:

[0101] Accuracy: The accuracy of data refers to the closeness between the collected or observed value of the data and the true value. The smaller the error, the higher the accuracy of the data.

[0102] Accuracy: The accuracy of data refers to the degree of similarity between different observations obtained when repeatedly measuring the same object. Accuracy is related to the accuracy of data collection. The higher the accuracy, the higher the accuracy of the data.

[0103] Consistency: Consistency refers to whether the data follows a unified specification and whether the data set maintains a unified format. Inconsistent data may lead to errors in data analysis.

[0104] Completeness: Completeness refers to whether data information is missing. Missing data will affect the accuracy and reliability of data analysis.

[0105] Timeliness: Timeliness refers to the time interval from when data is generated to when it can be viewed. Timely data is very important for decision support.

[0106] Validity: The value and format of the data must conform to the requirements of the data definition or business definition. For example, phone numbers and email addresses must conform to a specific format.

[0107] Uniqueness: The uniqueness of data means that there are no duplicate values ​​for a data item or a group of data. For example, ID data must be unique.

[0108] Security: Data security refers to whether data is protected during storage and transmission to prevent unauthorized access and tampering.

[0109] Accessibility: Data accessibility refers to the ease with which users can obtain and use data. Good accessibility can improve the efficiency of data use.

[0110] Authenticity: The authenticity of data refers to the high degree of controllability and traceability of the data collection process, ensuring the correctness of the data.

[0111] These indicators together form a comprehensive framework for data quality assessment, helping to ensure the reliability, consistency and availability of data. For different businesses, you can choose one or several of these indicators for assessment and decide whether to adopt them based on the assessment results.

[0112] S5: According to the needs of the business scenario and the evaluation status of each resource, select computing power points, storage points, algorithm models and business data to form a task combination to optimize the utilization and benefits of various resources.

[0113] Furthermore, the process of selecting computing power points, storage points, algorithm models and business data to form a task combination according to the needs of the business scenario and the evaluation status of each resource includes the following steps:

[0114] S51: Decompose each task according to the needs of the business scenario and clarify the specific task requirements; for example, see whether the task requirement is a classification problem, a regression problem, or other solution types.

[0115] S52: Select a suitable algorithm model based on the decomposed tasks and the data characteristics to be used. The appropriateness of the type of algorithm model determines whether the task can be completed. Therefore, it is preferred to select a suitable algorithm model based on the nature of the task.

[0116] S53: Select a computing power point according to the state of the algorithm model (for example, whether it is a state to be trained or a trained inference state to be used). In the management of the artificial intelligence platform, the calculation time of the computing power point often occupies most of the time of the entire business. Among them, the training time of the algorithm model is often long and the cost is high. Therefore, the computing power point is selected according to the state of the algorithm model. If it is a training task, then the computing power point with a large surplus of computing power is selected. The algorithm model that has been trained and is to be used for inference is arranged to the remaining computing power points, which can greatly reduce the completion time of the entire business.

[0117] S54: Select appropriate storage points and business data based on the available storage points and business data of the computing power point, taking into account factors such as bandwidth, distance, and data throughput.

[0118] First, determine the business data that can be used based on the nature of the task. For example, if the task is to predict the electricity consumption of a certain area, the available business data includes the historical electricity consumption data of the area and the weather, date and other influencing factors that affect electricity consumption. If the task is to predict the traffic volume of a certain section of road, the historical traffic volume of the section and the corresponding weather, date and other influencing factors data can be selected.

[0119] Secondly, determine the available storage points based on the scale of data calculation. For example, determine which storage points can meet the usage requirements based on the available storage capacity, bandwidth and other indicators of the storage points.

[0120] Finally, select specific storage points and business data based on distance, bandwidth speed, quantity, etc.

[0121] The specific selection calculation formula is as follows:

[0122]

[0123] In the above formula, X is the selection ratio; L is the distance between the storage point or business data and the computing power point; L P is the average distance between available storage points or business data and computing power points; D is the bandwidth between storage points or business data and computing power points; D P is the average bandwidth between available storage points or business data and computing power points; T is the data throughput of storage points or business data; T P It is the average throughput of storage points or business data. In some states, if a certain factor (such as jitter, packet loss) has a greater impact on data transmission, you can refer to the formula similar structure to add the calculation weight of the factor (the value of a certain point divided by all available average values) into it.

[0124] Furthermore, when selecting storage points, in order to consider the impact of storage margin on business expansion, that is, after the current business is completed, when the same business is started next time, there is a high probability that the stored data generated this time will be used, and new and more data will be generated at the same time, or as the calculation progresses, there is a certain probability that new data will be generated according to the progress. At this time, considering data access at the same storage point can improve the reliability of network data use. The specific calculation formula is as follows:

[0125]

[0126] In the above formula, X cSelect the weight for the storage point; L is the distance between the storage point or business data and the computing power point; L P is the average distance between available storage points or business data and computing power points; D is the bandwidth between storage points or business data and computing power points; D P is the average bandwidth between available storage points or business data and computing power points; T is the data throughput of storage points or business data; T P is the average throughput of storage points or business data; C is the storage space margin of the storage point; S is the amount of business data used in calculation; the ratio of C to S can show which storage point has a larger relative space margin.

[0127] In a specific implementation, the storage point or business data is selected based on the one with the largest weight.

[0128] The changes to the platform management architecture in this application plan are aimed at improving the efficiency of data processing. By driving computing through data flow, the underlying data is processed on demand and at the fastest speed, greatly improving data processing efficiency, thereby achieving a qualitative change in platform performance from quantitative change.

[0129] As production traffic exceeds consumption traffic in modern society, high-frequency and rich media are mainly based on instant interaction, which is sensitive to latency and has a large bandwidth demand. The modal development of generative AI will exceed Moore's Law in terms of network bandwidth demand, and the computing power layout will gradually form an architecture with centralized deployment of training at certain hub nodes and distributed deployment of reasoning. The centralized trained model needs to be synchronized to the distributed reasoning end, and the backbone network traffic is increasing, user data is isolated, and the metropolitan area traffic tide is obvious.

[0130] In this application scheme, computing power and network resources are integrated into a whole to form an efficient, scalable and flexible ubiquitous computing power network, so that the platform can evolve towards the integration of ubiquitous interconnection and ubiquitous computing unified scheduling and control, and realize the integrated platform service of local diversion and computing-network integration by driving the network to perceive the location of computing power.

[0131] By connecting ubiquitous computing power through the network, we can break through the performance limit of single-point computing power, give full play to the cluster advantages of computing power, and improve the scale efficiency of computing power. Through global intelligent scheduling and optimization of computing network resources, we can effectively promote the "flow" of computing power and meet the business needs for on-demand use of computing power.

[0132] The metering management of AI platform resources is the process of quantifying and evaluating the performance and effectiveness of AI platforms. The metering objects cover computing resources, storage resources, data set calls, model library calls, etc., to achieve user-level / code-level metering levels to achieve optimal utilization and benefits of IT resources (computing resources, storage resources, etc.) and atomic capabilities (data sets, model libraries, etc.). According to the service objects, it is divided into user-oriented core metering and platform operator-oriented core metering. The core metering topics for users include: historical cost measurement at the instance level or task level, and estimated measurement of future usage. The core metering topics for platform operators include: research on resource allocation, resource service consumption rankings, capacity water level warnings, etc.

[0133] Embodiment 2

[0134] like Figure 2 The figure shows a schematic diagram of the structure of an artificial intelligence platform resource metering management system, which is applied to the artificial intelligence platform resource metering management method described above, including:

[0135] The computing power point measurement module is used to detect and measure the computing power surplus of computing power points at various locations in the platform system, and divide the computing power points into training points and inference points according to the idle computing power surplus and computing power point type;

[0136] The storage point metering module is used to detect and measure the storage performance indicators of storage points at various locations in the platform system to obtain the parameters of each storage point;

[0137] Model measurement module, which is used to evaluate the algorithm model in the platform system and determine the type, performance indicators and status of the algorithm model;

[0138] Business data measurement module, which is used to detect and evaluate the business data type, nature and purpose, location distance or data volume within the platform system;

[0139] The task combination module is used to select computing power points, storage points, algorithm models and business data to form a task combination based on the needs of the business scenario and the evaluation status of each resource, so as to optimize the utilization and benefits of various resources.

[0140] Embodiment 3

[0141] A computer-readable storage medium, the computer-readable storage medium comprising a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the artificial intelligence platform resource metering management method described above.

[0142] Embodiment 4

[0143] A processor is used to run a program, wherein the program executes the artificial intelligence platform resource metering management method described above when running.

[0144] The present application provides a resource metering management method for an artificial intelligence platform, which specifically includes the following steps: S1: Detect and measure the surplus amount of computing power at each computing power point in the platform system, and divide the computing power points into training points and reasoning points according to the surplus amount of idle computing power and the type of computing power point; S2: Detect and measure the storage performance indicators of each storage point in the platform system to obtain the parameters of each storage point; S3: Evaluate the algorithm model in the platform system to determine the type, performance indicators and status of the algorithm model; S4: Detect and evaluate the business data type, nature and purpose, location distance or data volume in the platform system; S5: According to the needs of the business scenario and the evaluation status of each resource, select computing power points, storage points, algorithm models and business data to form a task combination, so that the utilization rate and benefit of various resources are optimized. In the present application scheme, computing power and network resources are integrated into a whole to form an efficient, scalable and flexible ubiquitous computing power network, so that the platform evolves towards the integration direction of ubiquitous interconnection and ubiquitous computing unified scheduling and unified control, and realizes the integrated platform service of nearby diversion and computing network integration by driving the network to perceive the location of computing power. By connecting ubiquitous computing power through the network, we can break through the performance limit of single-point computing power, give full play to the cluster advantages of computing power, and improve the scale efficiency of computing power. Through global intelligent scheduling and optimization of computing network resources, we can effectively promote the "flow" of computing power and meet the business needs for on-demand use of computing power.

[0145] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0146] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.

[0147] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0148] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nlyMemory), random access memory (RAM, RandomAccessMemory), mobile hard disk, magnetic disk or optical disk, etc., which can store program code.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A resource metering management method for an artificial intelligence platform, characterized in that: The specific steps include: S1: Detect and measure the surplus computing power of computing power points at various locations in the platform system, and divide the computing power points into training points and inference points according to the surplus idle computing power and computing power point types; S2: Detect and measure the storage performance indicators of each storage point in the platform system to obtain the parameters of each storage point; S3: Evaluate the algorithm model within the platform system and determine the type, performance indicators and status of the algorithm model; S4: Detect and evaluate the business data type, nature and purpose, location distance or data volume within the platform system; S5: According to the needs of the business scenario and the evaluation status of each resource, select computing power points, storage points, algorithm models and business data to form a task combination to optimize the utilization and benefits of various resources.

2. The artificial intelligence platform resource metering management method according to claim 1 is characterized in that: The computing power points include the same type of computing power points and heterogeneous computing power points; the computing power surplus detection and measurement formula of the same type of computing power points is as follows: A = a*b*(1-c); In the above formula, A is the surplus computing power of a certain type of processor; a is the performance of a single processor; b is the number of processors; and c is the utilization rate of each processor.

3. The artificial intelligence platform resource metering management method according to claim 2 is characterized in that: The calculation formula for detecting the surplus computing power of the heterogeneous computing power point is as follows: In the above formula, A i is the computing power surplus of the i-th type of processor; Y is the computing power surplus of the heterogeneous computing power point; n is the number of processor types in the heterogeneous computing power point.

4. The artificial intelligence platform resource metering management method according to claim 1 is characterized in that: The storage performance indicators of the storage point include IOPS, throughput, latency, bandwidth, data protection, and storage margin.

5. The artificial intelligence platform resource metering management method according to claim 1 is characterized in that: The process of selecting computing power points, storage points, algorithm models and business data to form a task combination according to the needs of the business scenario and the evaluation status of each resource includes the following steps: S51: Decompose the tasks according to the needs of the business scenario and clarify the specific task requirements; S52: Select a suitable algorithm model according to the decomposed tasks and the characteristics of the data to be used; S53: Select a computing power point according to the state of the algorithm model; S54: Select storage points and business data based on the available storage points and business data according to the computing power point, taking into account bandwidth, distance and data throughput factors.

6. The artificial intelligence platform resource metering management method according to claim 5 is characterized in that: Step S54 specifically includes the following steps: First, determine the business data that can be used based on the nature of the task; Secondly, determine the storage points that can be used based on the scale of data calculation; Finally, select specific storage points and business data based on distance, bandwidth speed, and quantity.

7. The artificial intelligence platform resource metering management method according to claim 6 is characterized in that: The calculation formula for selecting specific storage points and business data is as follows: In the above formula, X is the selection ratio; L is the distance between the storage point or business data and the computing power point; L P is the average distance between available storage points or business data and computing power points; D is the bandwidth between storage points or business data and computing power points; D P is the average bandwidth between available storage points or business data and computing power points; T is the data throughput of storage points or business data; T P It is the average throughput of storage points or business data.

8. An artificial intelligence platform resource metering management system, characterized in that: The artificial intelligence platform resource metering management method applied to any one of claims 1 to 7 comprises: The computing power point measurement module is used to detect and measure the computing power surplus of computing power points at various locations in the platform system, and divide the computing power points into training points and inference points according to the idle computing power surplus and computing power point type; The storage point metering module is used to detect and measure the storage performance indicators of storage points at various locations in the platform system to obtain the parameters of each storage point; Model measurement module, which is used to evaluate the algorithm model in the platform system and determine the type, performance indicators and status of the algorithm model; Business data measurement module, which is used to detect and evaluate the business data type, nature and purpose, location distance or data volume within the platform system; The task combination module is used to select computing power points, storage points, algorithm models and business data to form a task combination based on the needs of the business scenario and the evaluation status of each resource, so as to optimize the utilization and benefits of various resources.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the artificial intelligence platform resource metering management method according to any one of claims 1 to 7.

10. A processor, characterized in that: The processor is used to run a program, wherein the program, when running, executes the artificial intelligence platform resource metering management method described in any one of claims 1 to 7.

Images