Method and device for metering computing power utilization rate of intelligent computing center for common computing power

Abstract

The invention provides an intelligent computing center computing power utilization rate metering method and device oriented to common computing power, and relates to the technical field of intelligent computing centers, intelligent computing centers and computing power infrastructures, and the method comprises the steps: S1, obtaining the historical data of the utilization rate of a computing power cluster; s2, calculating second scheduling weights of the computing power resource nodes of the plurality of nodes based on computing power resource demand calculation results in a future time period and the first scheduling weights of the computing power resource nodes of the plurality of nodes; and S3, calculating the computing power resource utilization rate calculation information of the target GPU in the future time period based on the second scheduling weight of the computing power resource node and the feature information of the target GPU. According to the method, the calculation accuracy of the computing power resource utilization rate of the target GPU can be greatly improved, a large amount of computing power resources are prevented from being wasted, the economic cost is greatly reduced, and wide application of the general computing power is facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent computing centers, smart computing centers and computing power infrastructure, and in particular to a method and device for measuring computing power usage rate of an intelligent computing center for universal computing power. Background Art

[0002] With the rapid development of artificial intelligence technology, "intelligent computing centers" and "intelligent computing centers" have come into being.

[0003] "Intelligent computing center" refers to a facility that provides the required computing power, data and algorithms for artificial intelligence applications (such as artificial intelligence deep learning model development, model training and model reasoning) by using large-scale heterogeneous computing resources, including general computing power and intelligent computing power. The intelligent computing center covers facilities, hardware, and software, and can provide full-stack capabilities from bottom-level computing power to top-level application enablement.

[0004] “Intelligent Computing Center” includes but is not limited to “Smart Computing Center”.

[0005] "Intelligent Computing Center" refers to an artificial intelligence computing center. It is a type of computing infrastructure that is based on artificial intelligence theory, adopts artificial intelligence computing architecture, and provides computing power services, data services, and algorithm services required for artificial intelligence applications.

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

[0007] With the continuous development of artificial intelligence technology, graphics processing units (GPUs) have played an increasingly important role in people's lives. However, the computing resource utilization rates of different GPUs are different. Therefore, before using the GPU, it is necessary to calculate the computing resource utilization rate of the GPU. Currently, the computing resource utilization rate of the GPU usually needs to be calculated manually. It can be seen that the current calculation accuracy of the computing resource utilization rate of the GPU is very poor, resulting in a large waste of computing resources and high economic costs, which is not conducive to the widespread application of inclusive computing power. Summary of the invention

[0008] The present invention provides a method and device for measuring the computing power utilization rate of an intelligent computing center for inclusive computing power, which is used to solve the problem that the calculation accuracy of the computing power resource utilization rate of a GPU is very poor, resulting in a large amount of computing power resources being wasted, the economic cost is very high, and it is not conducive to the widespread application of inclusive computing power.

[0009] In order to solve the above-mentioned technical problems, the present invention is achieved as follows: In a first aspect, the present invention provides a method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power, comprising: Step S1: Obtaining historical usage data of a computing power cluster, where the historical usage data is historical data of computing power resources used by the computing power cluster; Step S2: based on the calculation result of the computing power resource demand in the future time period and the first scheduling weights of the computing power resource nodes of the multiple nodes, calculate the second scheduling weights of the computing power resource nodes of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the first scheduling weights of the computing power resource nodes are used to represent the historical scheduling weights of the computing power resources between the multiple nodes, the second scheduling weights of the computing power resource nodes are used to represent the scheduling weights of the computing power resources between the multiple nodes in the future time period, and the computing power resource demand calculation result is calculated based on the historical usage data of the computing power cluster; Step S3: Based on the second scheduling weight of the computing resource node and the characteristic information of the target GPU, calculate the computing resource utilization calculation information of the target GPU in the future time period, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

[0010] Optionally, the step S1 includes: Step S11: Obtain the historical usage data of the computing power cluster, and input the historical usage data into the Long Short-Term Memory (LSTM) model to calculate the computing power resource demand, and output the calculation result of the computing power resource demand of the computing power cluster in the future time period. The LSTM model applies an attention mechanism.

[0011] Optionally, step S2 includes: Step S21: Input the first scheduling weight of the computing power resource node and the calculation result of the computing power resource demand into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node of the multiple nodes. A multi-head attention mechanism is applied in the deep learning model.

[0012] Optionally, the step S21 includes: Step S211: constructing a node competition graph corresponding to the first scheduling weight of the computing power resource node, wherein the node competition graph is used to calculate the scheduling weights of the computing power resources between the multiple nodes; Step S212: Input the node competition graph and the computing power resource demand calculation result into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node.

[0013] Optionally, the S3 includes: Step S31: input the second scheduling weight of the computing power resource node and the characteristic information into the tree model to calculate the utilization rate of the computing power resources in the future time period, and output the computing power resource utilization rate calculation information.

[0014] Optionally, the feature information includes at least one of the following: static feature information, dynamic feature information, and context feature information of the target GPU; The static feature information is used to represent the fixed parameters of the target GPU, the dynamic feature information is used to represent the dynamically changing parameters of the target GPU during use, and the context feature information is used to represent the parameters of other GPUs connected to the same node as the target GPU.

[0015] Optionally, it also includes: Step S4: Calculate whether the computing power resource utilization calculation information is less than a preset threshold; Step S5: When the computing power resource utilization calculation information is less than the preset threshold, the computing power resource scheduling of the target GPU is controlled according to the computing power resource utilization calculation information.

[0016] Optionally, it also includes: Step S6: When the computing power resource utilization calculation information is greater than or equal to the preset threshold, an early warning signal is output, wherein the early warning signal is used to indicate that there is a deviation in the computing power resource utilization calculation information.

[0017] In a second aspect, the present invention provides a computing power utilization rate metering device for an intelligent computing center for universal computing power, comprising: An acquisition module is used to acquire historical usage data of a computing power cluster, where the historical usage data is historical data of computing power resources used by the computing power cluster; A first calculation module is used to calculate the second scheduling weights of the computing resource nodes of the multiple nodes based on the computing resource demand calculation result in the future time period and the computing resource node first scheduling weights of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the computing resource node first scheduling weight is used to represent the historical scheduling weights of the computing resources between the multiple nodes, the computing resource node second scheduling weight is used to represent the scheduling weights of the computing resources between the multiple nodes in the future time period, and the computing resource demand calculation result is calculated based on the usage history data of the computing power cluster; The second calculation module is used to calculate the computing resource utilization calculation information of the target GPU in the future time period based on the second scheduling weight of the computing resource node and the characteristic information of the target GPU, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

[0018] In a third aspect, the present invention provides an electronic device comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power as described in the first aspect above.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power as described in the first aspect above.

[0020] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power as described in the first aspect above.

[0021] In the present invention, historical utilization data of a computing power cluster is obtained, and the historical utilization data is historical data of computing power resources used by the computing power cluster; based on the calculation result of computing power resource demand in a future time period and the first scheduling weights of computing power resource nodes of multiple nodes, the second scheduling weights of computing power resource nodes of the multiple nodes are calculated, the multiple nodes are nodes in the computing power cluster, the first scheduling weights of the computing power resource nodes are used to represent the historical scheduling weights of computing power resources between the multiple nodes, the second scheduling weights of the computing power resource nodes are used to represent the scheduling weights of computing power resources between the multiple nodes in the future time period, and the computing power resource demand calculation result is calculated based on the historical utilization data of the computing power cluster; based on the second scheduling weight of the computing power resource node and characteristic information of a target graphics processor GPU, the computing power resource utilization calculation information of the target GPU in the future time period is calculated, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

[0022] In this way, the computing resource demand calculation result is first calculated through the utilization history data of the computing power cluster, and then the second scheduling weights of the computing resource nodes of multiple nodes are calculated through the computing resource demand calculation result and the first scheduling weights of the computing resource nodes of multiple nodes. Finally, based on the second scheduling weights of the computing resource nodes and the characteristic information of the target GPU, the computing resource utilization calculation information of the target GPU in the future time period is calculated. In this way, when calculating the computing resource utilization calculation information of the target GPU in the future time period, the utilization history data of the computing power cluster and the first scheduling weights of the computing resource nodes of multiple nodes are considered, and the utilization history data of the computing power cluster and the first scheduling weights of the computing resource nodes of multiple nodes are strongly correlated with the target GPU, thereby improving the accuracy of the computing resource utilization calculation information of the target GPU in the future time period obtained by the final calculation, avoiding a large amount of waste of computing power resources, greatly reducing economic costs, and facilitating the widespread application of inclusive computing power. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 One of the flow charts of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power provided by the present invention; Figure 2 The second flow chart of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power provided by the present invention; Figure 3 A schematic diagram of the structure of a computing power utilization rate metering device for an intelligent computing center oriented to inclusive computing power provided by the present invention; Figure 4 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the present invention to clearly and completely describe the technical solution in the present invention. Obviously, the described content is part of the content of the present invention, but not all of it. Based on the content in the present invention, all other contents obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] First, the technical terms involved in the present invention are briefly explained below.

[0026] The "computing power" mentioned in the present 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 execute certain computing requirements, and the computing power to achieve target result output by processing information data. It is a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity, and it mainly provides services to the society through computing power infrastructure.

[0027] The "computing power" (CP) described in the present invention refers to: it is the ability of a data center server to process data and output results. It is a comprehensive indicator to measure the computing power of a data center, including general computing power, super computing power and intelligent computing power. The commonly used unit of measurement is the number of floating-point operations performed per second (FLOPS, 1EFLOPS=10^18 FLOPS). The larger the value, the stronger the comprehensive computing power. According to calculations, 1 EFLOPS is about the computing power output of 5 Tianhe-2A or 500,000 mainstream server CPUs or 2 million mainstream notebooks. The calculation formula is: CP=CP General+CP Intelligent+CP Super.

[0028] The "carrying capacity" (Network Power, NP) mentioned in the present invention refers to: it is the performance of the data transmission capacity of the computing power facilities, including the comprehensive capabilities of network architecture, network bandwidth, transmission delay, intelligent management and scheduling, etc. It involves network transmission within and between data centers, and is a comprehensive indicator for measuring network transmission scheduling capabilities.

[0029] The "Storage Power" (SP) mentioned in the present invention refers to: the comprehensive capabilities of a data center in terms of data storage capacity, performance, safety and reliability, and green and low-carbon. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and server built-in storage devices. The commonly used unit of measurement for storage capacity is exabyte (EB, 1EB=2^60bytes), and the commonly used unit of measurement for performance is the number of reads and writes per second per unit capacity (IOPS / TB, Input / Output Operations Per Second / TB). The disaster recovery ratio is an important manifestation of safety and reliability.

[0030] The "computing power infrastructure" mentioned in the present invention refers to: a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, which can realize centralized computing, storage, transmission and application of information, and presents characteristics such as diversity and ubiquity, intelligence and agility, safety and reliability, green and low-carbon. It is of great significance to promote industrial transformation and upgrading, enable my country's scientific and technological innovation, meet people's better life and realize high-efficiency social governance.

[0031] The "new information infrastructure" mentioned in the present invention refers to: mainly including network infrastructure such as 5G network, fiber broadband network, backbone network, international communication network, satellite Internet, computing power infrastructure such as data center, general computing power center, intelligent computing center, supercomputing center, and new technology facilities such as artificial intelligence, blockchain, and quantum computing. With the emergence and promotion and application of new general technologies, the new information infrastructure will become more diverse.

[0032] The “computing power” mentioned in the present invention includes: general computing power, intelligent computing power and super computing power.

[0033] The “general computing power” mentioned in the present 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.

[0034] The "intelligent computing power" mentioned in the present invention refers to: a computing platform based on GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit) and other dedicated chips for various innovative artificial intelligence applications, such as natural language processing, machine vision, etc.

[0035] The "super computing power" mentioned in the present invention refers to: the computing power mainly provided by high-performance computing power clusters such as supercomputers. It uses 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 calculations in cutting-edge scientific fields, such as planetary simulation, drug molecule design, gene analysis, etc.

[0036] The "intelligent computing center" mentioned in the present invention refers to a facility that provides the required computing power, data and algorithms for artificial intelligence applications (such as artificial intelligence deep learning model development, model training and model reasoning) by using large-scale heterogeneous computing resources, including general computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.). The intelligent computing center covers facilities, hardware, and software, and can provide full-stack capabilities from bottom-level computing power to top-level application enablement.

[0037] The "intelligent computing center" mentioned in the present invention includes but is not limited to an intelligent computing center.

[0038] The "intelligent computing center" mentioned in the present invention is an artificial intelligence computing center, which is a type of computing power infrastructure that is based on artificial intelligence theory, adopts artificial intelligence computing architecture, and provides computing power services, data services, and algorithm services required for artificial intelligence applications.

[0039] The "computing power center" mentioned in the present invention refers to: a facility that is mainly composed of infrastructure such as wind, fire, water, electricity, and IT hardware and software equipment, and has computing power, transportation capacity, and storage capacity, including general data centers, intelligent computing centers, supercomputing centers, etc.

[0040] The "supercomputing center" mentioned in the present invention refers to: a supercomputing data center, which is a data center based on a supercomputer or a large-scale computing cluster, which can provide functions such as large-scale computing, storage and network services, and is widely used in application scenarios such as aerospace, national defense, oil exploration, climate modeling and genome sequencing.

[0041] The "computing resources" mentioned in the present invention refer to: technologies and facilities with information computing, transmission, storage and application capabilities required for the development of a digital society, 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 guarantee resources such as wind, fire, water and electricity.

[0042] The "inclusive computing power" mentioned in the present invention refers to providing appropriate and effective computing power services at an affordable cost to all social classes and groups that have computing power service needs based on the requirements of equal opportunity and the principle of commercial sustainability.

[0043] See also Figure 1 , Figure 1 : is a flow chart of a method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power provided by the present invention. Figure 1 As shown, the following steps are included: Step S1: Obtaining historical usage data of a computing power cluster, where the historical usage data is historical data of computing power resources used by the computing power cluster; It should be noted that the present invention can be applied to electronic devices, and the electronic devices may include a resource aggregation layer, or the electronic devices are connected to the resource aggregation layer, and the usage history data of the computing power cluster can be stored in the resource aggregation layer. In this way, the electronic devices can obtain the above-mentioned usage history data from the resource aggregation layer, or the above-mentioned usage history data can be stored in other electronic devices, and the other electronic devices may be servers, and the electronic devices in the present invention can obtain the above-mentioned usage history data from other electronic devices through the resource aggregation layer.

[0044] In addition, the triggering condition of step S1 may include at least one of the following: every preset period, and when the load fluctuation of the computing power cluster exceeds the load threshold, the specific values ​​of the above preset period and load threshold are not limited here.

[0045] Among them, the specific type of usage history data is not limited here. Optionally, the usage history data may include the usage history data of the GPU of the computing power cluster, that is, the computing power cluster may include multiple nodes, and each node may include multiple GPUs. The usage history data of the GPU of the computing power cluster may refer to the sum of the usage history data of at least some of the GPUs included in the above-mentioned computing power cluster.

[0046] Optionally, the usage history data may also include at least one of the following: the central processing unit (CPU) usage history data of the computing power cluster, the memory (Memory) usage history data of the computing power cluster, and the input / output (IO) interface usage history data. It should be noted that the above-mentioned CPU usage history data, memory usage history data and IO interface usage history data are all associated with the GPU usage history data of the computing power cluster. For example, the higher the usage rate of the GPU of the computing power cluster, the higher the CPU usage rate, memory usage rate and IO interface usage rate. In this way, the richer the content included in the usage history data, the higher the accuracy of the computing resource usage calculation information of the calculated target GPU in the future time period.

[0047] It should be noted that the time period corresponding to the historical data of the utilization rate of the computing power cluster is not limited here. Optionally, the time period corresponding to the historical data of the utilization rate of the computing power cluster can be a first time period before the current moment. The start time and end time of the first time period can be the first moment and the second moment respectively. The interval between the second moment and the current moment can be a preset time period, and the specific value of the preset time period is not limited here. Optionally, the preset time period can be 0, that is, the second moment is the current moment; optionally, the preset time period can be greater than 0, that is, the second moment is not the current moment.

[0048] Step S2: based on the calculation result of the computing power resource demand in the future time period and the first scheduling weights of the computing power resource nodes of the multiple nodes, calculate the second scheduling weights of the computing power resource nodes of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the first scheduling weights of the computing power resource nodes are used to represent the historical scheduling weights of the computing power resources between the multiple nodes, the second scheduling weights of the computing power resource nodes are used to represent the scheduling weights of the computing power resources between the multiple nodes in the future time period, and the computing power resource demand calculation result is calculated based on the historical usage data of the computing power cluster; The specific value of the future time period is not limited here. Optionally, the future time period can be a first cycle starting from the current moment and lasting for 30 minutes. Optionally, the future time period can also be a second cycle after the first cycle. The first cycle and the second cycle can have the same duration, and there can be a time interval between the first cycle and the second cycle. That is, the future time period can be understood as a future time period after the current moment.

[0049] Among them, the calculation result of computing power resource demand can also be referred to as the overall resource demand of the computing power cluster in the future time period.

[0050] Among them, the first scheduling weight of the computing power resource node can be understood as being used to represent the scheduling weight of computing power resources between multiple nodes. If the scheduling weight is larger, the node's utilization rate of computing power resources will be higher. For example: the first scheduling weight of the computing power resource node is used to represent the scheduling weight of different nodes for the CPU, or the first scheduling weight of the computing power resource node is used to represent the scheduling weight of different nodes for the memory bandwidth.

[0051] The storage form of the first scheduling weight of the computing power resource node is not limited here. Optionally, the first scheduling weight of the computing power resource node can be stored in a matrix or a curve graph. When the first scheduling weight of the computing power resource node is stored in a matrix form, the above matrix can be called a computing power resource interaction matrix; optionally, when the first scheduling weight of the computing power resource node is stored in a curve graph, the above curve graph can be called a computing power resource interaction curve graph.

[0052] Among them, the first scheduling weight of the computing power resource node can be understood as the scheduling weight of computing power resources between multiple nodes before the current moment, and the second scheduling weight of the computing power resource node can be understood as the scheduling weight of computing power resources between multiple nodes after the current moment, that is, the second scheduling weight of the computing power resource node can be understood as the scheduling weight of computing power resources between multiple nodes in the future time period.

[0053] Step S3: Based on the second scheduling weight of the computing resource node and the characteristic information of the target GPU, calculate the computing resource utilization calculation information of the target GPU in the future time period, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

[0054] Among them, the feature information of the target GPU can also be referred to as hardware feature information or feature engineering information of the target GPU, and the above-mentioned feature engineering information can be referred to as feature engineering, and the feature information of the target GPU can be calculated based on the parameters of the target GPU and / or the parameters of other GPUs connected to the same node as the target GPU.

[0055] Optionally, the target GPU may be one of the multiple GPUs corresponding to one of the multiple nodes; alternatively, the target GPU may be some of the multiple GPUs corresponding to one of the multiple nodes, which is not specifically limited here.

[0056] In the present invention, the computing power resource demand calculation result is first calculated through the utilization history data of the computing power cluster, and then the second scheduling weights of the computing power resource nodes of multiple nodes are calculated through the computing power resource demand calculation result and the first scheduling weights of the computing power resource nodes of multiple nodes. Finally, based on the second scheduling weights of the computing power resource nodes and the characteristic information of the target GPU, the computing power resource utilization calculation information of the target GPU in the future time period is calculated. In this way, when calculating the computing power resource utilization calculation information of the target GPU in the future time period, the utilization history data of the computing power cluster and the first scheduling weights of the computing power resource nodes of multiple nodes are considered, and the utilization history data of the computing power cluster and the first scheduling weights of the computing power resource nodes of multiple nodes are strongly correlated with the target GPU, thereby improving the accuracy of the computing power resource utilization calculation information of the target GPU in the future time period obtained by the final calculation, avoiding a large amount of waste of computing power resources, greatly reducing economic costs, and facilitating the widespread application of inclusive computing power.

[0057] Optionally, the step S1 includes: Step S11: Obtain the historical usage data of the computing power cluster, input the historical usage data into the LSTM model to calculate the computing power resource demand, and output the calculation result of the computing power resource demand of the computing power cluster in the future time period. The Attention mechanism is applied in the LSTM model.

[0058] Among them, since the time period corresponding to the usage history data can be the first time period before the current moment, that is, the first time period is different, the usage history data may also be different, that is, the usage history data is related to time, so the usage history data can be called time series data.

[0059] Optionally, step S11 includes: when the change value of the computing power resource demand calculation result is within a preset range or less than a preset change threshold, inputting the usage history data into the LSTM model to perform computing power resource demand calculation, and outputting the computing power resource demand calculation result of the computing power cluster in the future time period. The change value of the computing power resource demand calculation result is within a preset range or less than a preset change threshold, which can also be referred to as the computing power resource demand calculation result being stable.

[0060] In addition, when the change value of the computing power resource demand calculation result is not within the preset range or is greater than or equal to the preset change threshold, it can be said that the computing power resource demand calculation result is unstable. At this time, the LSTM model can be retrained.

[0061] In the present invention, since the LSTM model is good at processing time series data and can learn the periodic laws of changes in computing power resources through historical usage data, and the Attention mechanism can identify key time points and dynamically adjust the calculation weights of various input data, thereby improving the accuracy of the final calculated computing power resource demand calculation results, the above-mentioned input data can refer to different data in the historical usage data of the computing power cluster, for example: it can refer to data at different times.

[0062] Among them, the specific types of periodic patterns and key time points are not limited here. Optionally, the periodic pattern may refer to the difference pattern of peak computing load in the morning and evening, and the key time point may refer to the period of sudden traffic, such as the period when the traffic suddenly increases or the period when the traffic suddenly decreases.

[0063] It should be noted that since the consumption of computing power resources of the computing power cluster as a whole is highly correlated with the key fluctuation period, the configuration of the computing power resources of the computing power cluster as a whole needs to ensure that the needs of the key fluctuation period are met. For example, the configuration of the computing power resources of the computing power cluster as a whole needs to ensure that the needs of key switching periods such as the period when the traffic suddenly increases and the period when the traffic suddenly decreases are met at the same time. In this way, if the configuration of the overall resources of the computing power cluster is always configured according to the needs of the period when the traffic suddenly increases, it is easy to cause over-allocation of computing power resources in other periods, resulting in waste of resources. The present invention obtains the computing power resource demand calculation result of the computing power cluster in the future time period by calculating, so as to accurately calculate the demand for computing power resources at different times in the future time period, and flexibly configure the computing power resources at different times according to the demand, thereby avoiding waste of computing power resources.

[0064] For example, the application scenario of the present invention is to calculate the demand for computing resources during a period of e-commerce promotion, and expand the resource pool of the computing cluster during the period in advance before the e-commerce promotion according to the demand of the period.

[0065] Optionally, step S2 includes: Step S21: Input the first scheduling weight of the computing power resource node and the calculation result of the computing power resource demand into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node of the multiple nodes. A multi-head attention mechanism is applied in the deep learning model.

[0066] The specific type of the deep learning model is not limited here. Optionally, the deep learning model can be a Transformer model.

[0067] Among them, the second scheduling weight of the computing power resource node can be the scheduling weight of different nodes for computing power resources in the future time period.

[0068] In the present invention, a multi-head attention mechanism is applied in the deep learning model, and the deep learning model can separate resource competition modes of different dimensions through the multi-head attention mechanism, for example: separating CPU-intensive tasks and IO-intensive tasks, and outputting the second scheduling weight of the computing power resource node of each node corresponding to the different resource competition modes. In this way, the calculated second scheduling weight of the computing power resource node is more accurate, and the calculation method of the second scheduling weight of the computing power resource node is more flexible, thereby optimizing the calculation method of the second scheduling weight of the computing power resource node.

[0069] It should be noted that the vector corresponding to the first scheduling weight of the computing power resource node and the vector corresponding to the calculation result of the computing power resource demand can be extracted first, and then the vector corresponding to the first scheduling weight of the computing power resource node and the vector corresponding to the calculation result of the computing power resource demand can be input into the deep learning model for weight calculation, and the second scheduling weights of the computing power resource nodes of multiple nodes can be output.

[0070] For example: when the vector corresponding to the calculation result of the computing power resource demand can be used as the reference vector for position encoding, the vector corresponding to the first scheduling weight of the computing power resource node and the above-mentioned reference vector are input into the deep learning model for weight calculation, thereby outputting the second scheduling weights of the computing power resource nodes of multiple nodes.

[0071] Optionally, the step S21 includes: Step S211: constructing a node competition graph corresponding to the first scheduling weight of the computing power resource node, wherein the node competition graph is used to calculate the scheduling weights of the computing power resources between the multiple nodes; Step S212: Input the node competition graph and the computing power resource demand calculation result into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node.

[0072] Among them, by constructing a node competition graph, the node competition graph is used to calculate the scheduling weights of computing resources between multiple nodes. Different nodes have different scheduling weights for computing resources. In this way, the node competition graph can intuitively reflect the scheduling relationship of computing resources across nodes in multiple nodes.

[0073] In the present invention, by constructing a node competition graph, the scheduling weights of computing resources between multiple nodes can be calculated to avoid the emergence of "hungry nodes", that is, the scheduling of computing resources of multiple nodes is more balanced, and the second scheduling weights of computing resource nodes are optimized. At the same time, the weight calculation is performed through a deep learning model, and the second scheduling weights of computing resource nodes of multiple nodes are output. In this way, the calculation efficiency of the second scheduling weights of computing resource nodes can be improved, and the calculation accuracy of the second scheduling weights of computing resource nodes can be improved.

[0074] It should be noted that the above-mentioned hungry nodes can be understood as nodes with less computing power resources scheduled.

[0075] The application scenario of the present invention may be: when the computing power cluster is an open source container orchestration and management platform computing power cluster, and optionally, when the computing power cluster is a Kubernetes computing power cluster, during the scheduling process of the Kubernetes computing power cluster, the second scheduling weights of the computing power resource nodes of multiple nodes may be output, so that the load of the nodes corresponding to the GPU and the nodes corresponding to the non-GPU may be dynamically balanced.

[0076] Optionally, the S3 includes: Step S31: input the second scheduling weight of the computing power resource node and the characteristic information into the tree model to calculate the utilization rate of the computing power resources in the future time period, and output the computing power resource utilization rate calculation information.

[0077] The specific type of the tree model is not limited herein. Optionally, the tree model may include at least one of a lightweight gradient boosting machine (Light Gradient Boosting Machine, LightGBM) model and an extreme gradient boosting (eXtremeGradient Boosting, XGBoost) model.

[0078] In the present invention, the utilization rate of computing power resources in the future time period is calculated by a tree model, thereby outputting computing power resource utilization calculation information, and the second scheduling weight and feature information of the computing power resource node may include nonlinear features, and the tree model is good at processing the interaction of nonlinear features, so that the calculation efficiency of the computing power resource utilization calculation information can be improved; at the same time, the tree model can break through the feature importance analysis to identify key bottlenecks, for example: it can reduce the impact of the fragmentation index information included in the feature information on the inference delay, thereby further improving the calculation efficiency of the computing power resource utilization calculation information, and adjust or guide the target GPU and other GPUs through the output computing power resource utilization calculation information. The above-mentioned video memory can be understood as the computing power resources of the GPU.

[0079] When the feature information includes fragmentation index information and a batch size parameter, the tree model can be used to process the correlation between the fragmentation index information and the batch size parameter.

[0080] It should be noted that the application scenarios of the present invention may include: the tree model can be applied to deep learning training, so as to optimize the usage of video memory, thereby reducing the occurrence of memory overflow (Out Of Memory, OOM) errors. And the present invention can realize the calculation of fine-grained computing resources such as computing resource usage calculation information through the tree model and high-dimensional heterogeneous data (such as fragmentation index information).

[0081] Optionally, the feature information includes at least one of the following: static feature information, dynamic feature information, and context feature information of the target GPU; The static feature information is used to represent the fixed parameters of the target GPU, the dynamic feature information is used to represent the dynamically changing parameters of the target GPU during use, and the context feature information is used to represent the parameters of other GPUs connected to the same node as the target GPU.

[0082] The static feature information may include at least one of the following parameters of the target GPU: GPU model, video memory capacity, high-speed serial computer expansion bus standard (Peripheral Component Interconnect Express, PCIe) bandwidth, etc.

[0083] The dynamic feature information may include a video memory fragmentation index of the target GPU, and the video memory fragmentation index may also be referred to as a video memory fragmentation index, fragmentation index information, or a video memory fragmentation degree parameter.

[0084] It should be noted that the method for obtaining the above-mentioned video memory fragmentation index is not limited here. Optionally, the above-mentioned video memory fragmentation index can be obtained by real-time sampling through a GPU performance monitoring tool.

[0085] The context feature information may include at least one of the following: various parameters of other GPUs connected to the same node as the target GPU (such as the second scheduling weight of the computing power resource node) and process-level resource lock status information.

[0086] In the present invention, the feature information includes at least one of the following: static feature information, dynamic feature information and context feature information of the target GPU. In this way, the diversity of the types of feature information is increased. When the types of feature information are more, the accuracy of the computing resource utilization calculation information of the target GPU in the future time period obtained by the final calculation is also higher.

[0087] Optionally, it also includes: Step S4: Calculate whether the computing power resource utilization calculation information is less than a preset threshold; Step S5: When the computing power resource utilization calculation information is less than the preset threshold, the computing power resource scheduling of the target GPU is controlled according to the computing power resource utilization calculation information.

[0088] The specific value of the preset threshold is not limited here. Optionally, the preset threshold can be an empirical value obtained through multiple statistics during the training process. Optionally, the preset threshold can be a value downloaded from other electronic devices.

[0089] In the present invention, when the computing power resource utilization calculation information is less than the preset threshold, it means that the calculated computing power resource utilization calculation information has a high accuracy. Therefore, the computing power resource scheduling of the target GPU can be controlled according to the computing power resource utilization calculation information, thereby ensuring the accuracy of the computing power resource scheduling of the target GPU.

[0090] Optionally, it also includes: Step S6: When the computing power resource utilization calculation information is greater than or equal to the preset threshold, an early warning signal is output, wherein the early warning signal is used to indicate that there is a deviation in the computing power resource utilization calculation information.

[0091] In the present invention, when the computing power resource utilization calculation information is greater than or equal to the preset threshold, it means that there is a deviation in the calculated computing power resource utilization calculation information and there is a memory leakage phenomenon. Therefore, a warning signal can be output to enhance the warning effect.

[0092] It should be noted that when a warning signal is output, it means that there is a deviation in the calculation process of the computing power resource utilization information. At this time, the model used in the calculation process of the computing power resource utilization information can be calibrated to improve the calculation accuracy of the computing power resource utilization information. The above process can also be called a deviation back propagation mechanism or an error back propagation mechanism.

[0093] For example: when the computing resource utilization calculation information is greater than or equal to the preset threshold, the calibration of the computing model of the node corresponding to the target GPU can be triggered. When the cumulative error of the computing model corresponding to the node reaches the target value, the attention mechanism weight of the computing model of the computing cluster corresponding to the node can be updated. In this way, the end-to-end connectivity of the hierarchical computing of the target GPU, nodes and computing clusters can be achieved to ensure the consistency of the computing results. Optionally, the computing model of the computing cluster can include an LSTM model, and the computing model corresponding to the node can include a deep learning model.

[0094] See also Figure 2 , Figure 2A specific description is provided for the present invention to fully illustrate the above content, and Figure 2 The various technical features in can refer to the above related descriptions, and have the same beneficial technical effects, and will not be described in detail here. Optionally, Figure 2 The Transformer encoder in can be understood as the above-mentioned Transformer model, and the node resource weight can be understood as the second scheduling weight of the computing resource node of the above-mentioned multiple nodes, and the GPU load prediction can be understood as the calculation process of the computing resource utilization rate calculation information of the target GPU in the future time period, and the anomaly detection can be understood as the process of calculating whether the computing resource utilization rate calculation information is less than the preset threshold.

[0095] See also Figure 3 , Figure 3 A schematic diagram of a computing power utilization rate metering device for an intelligent computing center for universal computing power provided by the present invention is shown in FIG. Figure 3 As shown, the computing power utilization rate metering device 300 for intelligent computing center for universal computing power includes: The acquisition module 301 is used to acquire the usage history data of the computing power cluster, where the usage history data is the history data of the computing power cluster using the computing power resources; A first calculation module 302 is used to calculate the second scheduling weights of the computing resource nodes of the multiple nodes based on the computing resource demand calculation result in the future time period and the computing resource node first scheduling weights of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the computing resource node first scheduling weight is used to represent the historical scheduling weights of the computing resources between the multiple nodes, the computing resource node second scheduling weight is used to represent the scheduling weights of the computing resources between the multiple nodes in the future time period, and the computing resource demand calculation result is calculated based on the usage history data of the computing power cluster; The second calculation module 303 is used to calculate the computing resource utilization calculation information of the target GPU in the future time period based on the second scheduling weight of the computing resource node and the characteristic information of the target GPU, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

[0096] Optionally, the acquisition module 301 is also used to obtain the usage history data of the computing power cluster, and input the usage history data into the long short-term memory network LSTM model to calculate the computing power resource demand, and output the calculation result of the computing power resource demand of the computing power cluster in the future time period. The LSTM model applies an attention mechanism.

[0097] Optionally, the first calculation module 302 is also used to input the first scheduling weight of the computing power resource node and the calculation result of the computing power resource demand into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node of the multiple nodes, and a multi-head attention mechanism is applied in the deep learning model.

[0098] Optionally, the first calculation module 302 includes: A construction submodule, used to construct a node competition graph corresponding to the first scheduling weight of the computing power resource node, wherein the node competition graph is used to calculate the scheduling weight of the computing power resources between the multiple nodes; A calculation submodule is used to input the node competition graph and the calculation result of the computing power resource demand into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node.

[0099] Optionally, the second calculation module 303 is further used to input the second scheduling weight of the computing resource node and the characteristic information into the tree model to calculate the utilization rate of the computing resource in the future time period, and output the computing resource utilization rate calculation information.

[0100] Optionally, the feature information includes at least one of the following: static feature information, dynamic feature information, and context feature information of the target GPU; The static feature information is used to represent the fixed parameters of the target GPU, the dynamic feature information is used to represent the dynamically changing parameters of the target GPU during use, and the context feature information is used to represent the parameters of other GPUs connected to the same node as the target GPU.

[0101] Optionally, the computing power utilization rate metering device 300 for the intelligent computing center for universal computing power further includes: A third calculation module is used to calculate whether the computing power resource utilization calculation information is less than a preset threshold; The control module is used to control the computing resource scheduling of the target GPU according to the computing resource utilization calculation information when the computing resource utilization calculation information is less than the preset threshold.

[0102] Optionally, the computing power utilization rate metering device 300 for the intelligent computing center for universal computing power further includes: The output module is used to output a warning signal when the computing power resource utilization calculation information is greater than or equal to the preset threshold value, and the warning signal is used to indicate that there is a deviation in the computing power resource utilization calculation information.

[0103] The device 300 for measuring the computing power usage rate of an intelligent computing center for inclusive computing power provided by the present invention can execute each step of the method for measuring the computing power usage rate of an intelligent computing center for inclusive computing power, and thus has the same beneficial technical effects as the method for measuring the computing power usage rate of an intelligent computing center for inclusive computing power, and the details will not be repeated here.

[0104] Please refer to Figure 4 The present invention also provides an electronic device 40, including a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor 41. When the computer program is executed by the processor 41, the various processes shown in the above-mentioned method for measuring the computing power utilization rate of an intelligent computing center for inclusive computing power are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0105] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, each process shown in the above-mentioned intelligent computing center computing power utilization rate measurement method for inclusive computing power is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0106] The present invention also provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the above Figure 1 The various processes shown in the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power can achieve the same technical effect. To avoid repetition, they will not be repeated here.

[0107] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0108] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method provided by the above invention can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an 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, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the various methods provided by the present invention.

[0109] The present invention is described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for measuring computing power utilization rate of intelligent computing centers for universal computing power, characterized in that: include: Step S1: Obtaining historical usage data of a computing power cluster, where the historical usage data is historical data of computing power resources used by the computing power cluster; Step S2: based on the calculation result of the computing power resource demand in the future time period and the first scheduling weights of the computing power resource nodes of the multiple nodes, calculate the second scheduling weights of the computing power resource nodes of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the first scheduling weights of the computing power resource nodes are used to represent the historical scheduling weights of the computing power resources between the multiple nodes, the second scheduling weights of the computing power resource nodes are used to represent the scheduling weights of the computing power resources between the multiple nodes in the future time period, and the computing power resource demand calculation result is calculated based on the historical usage data of the computing power cluster; Step S3: Based on the second scheduling weight of the computing resource node and the characteristic information of the target graphics processor GPU, calculate the computing resource utilization calculation information of the target GPU in the future time period, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

2. The method according to claim 1, characterized in that The step S1 comprises: Step S11: Obtain the historical usage data of the computing power cluster, and input the historical usage data into the long short-term memory network LSTM model to calculate the computing power resource demand, and output the calculation result of the computing power resource demand of the computing power cluster in the future time period. The LSTM model applies an attention mechanism.

3. The method according to claim 1, characterized in that The step S2 comprises: Step S21: Input the first scheduling weight of the computing power resource node and the calculation result of the computing power resource demand into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node of the multiple nodes. A multi-head attention mechanism is applied in the deep learning model.

4. The method according to claim 3, characterized in that The step S21 comprises: Step S211: constructing a node competition graph corresponding to the first scheduling weight of the computing power resource node, wherein the node competition graph is used to calculate the scheduling weights of the computing power resources between the multiple nodes; Step S212: Input the node competition graph and the computing power resource demand calculation result into the deep learning model for weight calculation, and output the second scheduling weight of the computing power resource node.

5. The method according to claim 1, characterized in that The S3 includes: Step S31: input the second scheduling weight of the computing power resource node and the characteristic information into the tree model to calculate the utilization rate of the computing power resources in the future time period, and output the computing power resource utilization rate calculation information.

6. The method according to claim 5, characterized in that The feature information includes at least one of the following: static feature information, dynamic feature information, and context feature information of the target GPU; The static feature information is used to represent the fixed parameters of the target GPU, the dynamic feature information is used to represent the dynamically changing parameters of the target GPU during use, and the context feature information is used to represent the parameters of other GPUs connected to the same node as the target GPU.

7. The method according to any one of claims 1 to 6, characterized in that Also includes: Step S4: Calculate whether the computing power resource utilization calculation information is less than a preset threshold; Step S5: When the computing power resource utilization calculation information is less than the preset threshold, the computing power resource scheduling of the target GPU is controlled according to the computing power resource utilization calculation information.

8. The method according to claim 7, characterized in that Also includes: Step S6: When the computing power resource utilization calculation information is greater than or equal to the preset threshold, an early warning signal is output, wherein the early warning signal is used to indicate that there is a deviation in the computing power resource utilization calculation information.

9. A computing power utilization rate metering device for intelligent computing centers oriented to universal computing power, characterized in that: include: An acquisition module is used to acquire historical usage data of a computing power cluster, where the historical usage data is historical data of computing power resources used by the computing power cluster; A first calculation module is used to calculate the second scheduling weights of the computing resource nodes of the multiple nodes based on the computing resource demand calculation result in the future time period and the computing resource node first scheduling weights of the multiple nodes, the multiple nodes are nodes in the computing power cluster, the computing resource node first scheduling weight is used to represent the historical scheduling weights of the computing resources between the multiple nodes, the computing resource node second scheduling weight is used to represent the scheduling weights of the computing resources between the multiple nodes in the future time period, and the computing resource demand calculation result is calculated based on the usage history data of the computing power cluster; The second calculation module is used to calculate the computing resource utilization calculation information of the target GPU in the future time period based on the second scheduling weight of the computing resource node and the characteristic information of the target GPU, and the target GPU is the GPU corresponding to at least some of the multiple nodes.

10. 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, the steps of the method for measuring computing power usage rate of an intelligent computing center for inclusive computing power as described in any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for measuring computing power usage rate of an intelligent computing center for inclusive computing power as described in any one of claims 1 to 8.

12. A computer program product, characterized in that It includes computer instructions, which, when executed by a processor, implement the steps of the method for measuring computing power utilization rate of an intelligent computing center for inclusive computing power as described in any one of claims 1 to 8.

Images