Computing power scheduling method and device of intelligent computing center

By analyzing the operator execution flowchart of the algorithm and allocating computing resources as needed, the problem of waste of computing resources in the intelligent computing center is solved, and efficient resource utilization and performance improvement is achieved.

CN119065816APending Publication Date: 2024-12-03DATACANVAS LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411260082.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In the intelligent computing center, the use of a single hardware configuration or runtime environment to process computing steps leads to insufficient utilization of computing power resources, resulting in waste of resources, and limits the improvement of computing power.

Method used

By obtaining the operator execution flowchart of the algorithm, analyzing and determining the execution order of the operator unit, allocating the computing power resources required for each operator unit based on the preset operator library, avoiding the use of a single hardware configuration to process all steps, and realizing each Each operator unit allocates resources as needed.

Benefits of technology

Accurate control and efficient use of computing resources have improved the computing resource utilization rate of the intelligent computing center, improved computing performance, and avoided resource waste.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119065816A_ABST
    Figure CN119065816A_ABST
Patent Text Reader

Abstract

The invention provides a computing power scheduling method and device of an intelligent computing center. The method comprises the steps that S1, an operator execution flow chart of an algorithm of a problem to be solved is acquired; the operator execution flow chart comprises a plurality of operator units and connecting lines used for connecting the operator units, and the output of the operator unit corresponding to the starting point of the connecting lines serves as the input of the operator unit corresponding to the end point of the connecting lines; s2, analyzing the operator execution flow chart to determine an execution sequence of all operator units; s3, based on a preset operator library, computing power resources corresponding to each operator unit in all the operator units are determined, and S4, the corresponding computing power resources are distributed to each operator unit, and each operator unit is executed according to the execution sequence. Therefore, only each operator is allocated with the actually required computing power resource quantity, the computing power resource utilization rate of the intelligent computing center is improved, the computing power of the intelligent computing center is further improved, and high-performance operation of the intelligent computing center is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of computing power technology, and in particular, to a computing power scheduling method and device for an intelligent computing center. Background Art

[0002] With the development of artificial intelligence technology and computing power technology, the concept of intelligent computing center has emerged. "Intelligent computing center" refers to the use of large-scale heterogeneous computing power resources, including general computing power and intelligent computing power, mainly for artificial intelligence applications (such as artificial intelligence deep learning model development, model training and model reasoning and other scenarios) to provide the required computing power, data and algorithms. Intelligent computing center covers facilities, hardware, software, and can provide full-stack capabilities from bottom-level computing power to top-level application enablement.

[0003] In the operation of intelligent computing centers, a complete computer problem usually consists of several algorithmic steps, and the computing resources required for different algorithmic steps often vary greatly. For example, a typical computer problem, such as data analysis, machine learning, or scientific simulation, usually involves multiple stages, including data preprocessing, core computing, and result analysis, and the computing power requirements for each stage vary significantly. In the data preprocessing stage, it mainly involves data cleaning and organization. These tasks usually do not require too many CPU (Central Processing Unit) or GPU (Graphics Processing Unit) resources, but in the core computing stage, such as the training of machine learning models or complex numerical simulations, a large amount of high-performance computing resources may be required.

[0004] However, in the current computer problem solving in intelligent computing centers, a single hardware configuration or runtime environment is usually used to handle all computing steps, which often leads to the underutilization and waste of computing resources, thus limiting the computing power improvement of intelligent computing centers. Summary of the invention

[0005] The embodiments of the present invention provide a computing power scheduling method and device for an intelligent computing center to solve the technical problem that in the current computer problem solving in the intelligent computing center, a single hardware configuration or runtime environment is usually used to process all computing steps, resulting in insufficient utilization of computing power resources, causing a waste of computing power resources, and thus limiting the improvement of the computing power of the intelligent computing center.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0007] In a first aspect, an embodiment of the present invention provides a computing power scheduling method for an intelligent computing center, the method comprising:

[0008] Step S1: obtaining an operator execution flow chart of an algorithm for a problem to be solved; wherein the operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line serves as the input of the operator unit corresponding to the end point of the line;

[0009] Step S2: parsing the operator execution flow chart to determine the execution order of all operator units;

[0010] Step S3: Based on the preset operator library, the computing resources corresponding to each operator unit in all the operator units are determined respectively.

[0011] Step S4: Allocate corresponding computing resources to each operator unit respectively, and execute each operator unit in the execution order.

[0012] Optionally, the step S1 includes:

[0013] Step S11: calling a canvas tool to create a canvas;

[0014] Step S12: receiving an operation instruction from a user, and determining the operator execution flowchart on the canvas according to the operation instruction, wherein the operation instruction includes: an operation of selecting an operator unit from the preset operator library and an operation of connecting lines between the operator units.

[0015] Optionally, step S3 includes:

[0016] Step S31: Based on the preset operator library, metadata information corresponding to each operator unit is determined respectively;

[0017] Step S32: Based on the metadata information, determine the computing power resources corresponding to each operator unit respectively, wherein the metadata information includes at least one of the following: the name of the operator unit, the algorithm introduction corresponding to the operator unit, the computing power resources corresponding to the operator unit, the input parameters corresponding to the operator unit, the input data set of the operator unit, the output data set of the operator unit, and the output parameters of the operator unit.

[0018] Optionally, step S4 includes:

[0019] Step S41: parameterizing the computing power resources to obtain parameterized computing power resources;

[0020] Step S42: Based on the parameterized computing resources, call the container orchestration management tool to allocate corresponding computing resources to each operator unit, and execute each operator unit in the execution order.

[0021] Optionally, the step S42 includes:

[0022] Step S421: constructing a configuration file of the container orchestration management tool based on the parameterized computing resources;

[0023] Step S422: Creating job resources of the container orchestration management tool according to the configuration file and the parameterized computing resources;

[0024] Step S423: calling the computing power resources according to the job resources, and executing each operator unit in the execution order.

[0025] Optionally, step S4 further includes:

[0026] Step S43: after the current operator unit is executed, generating an execution status corresponding to the current operator unit, wherein the execution status includes any one of the following: successful execution, failed execution, and executing;

[0027] Step S44: Determine the next operator unit to be executed according to the execution state and the execution order.

[0028] In a second aspect, an embodiment of the present invention provides a computing power scheduling device for an intelligent computing center, the device comprising:

[0029] An acquisition module is used to acquire an operator execution flow chart of an algorithm for a problem to be solved; wherein the operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line serves as the input of the operator unit corresponding to the end point of the line;

[0030] An execution module is used to parse the operator execution flow chart to determine the execution order of all operator units; based on a preset operator library, the computing power resources corresponding to each operator unit in all the operator units are determined respectively, the corresponding computing power resources are allocated to each operator unit respectively, and each operator unit is executed in the execution order.

[0031] In a third aspect, an embodiment of 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 when the program is executed by the processor, the steps of a method for scheduling computing power in an intelligent computing center as described in the first aspect are implemented.

[0032] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of a computing power scheduling method for an intelligent computing center as described in the first aspect are implemented.

[0033] In a fifth aspect, an embodiment of the present invention provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of a computing power scheduling method for an intelligent computing center as described in the first aspect.

[0034] In an embodiment of the present invention, an operator execution flowchart is first created to provide a clear and intuitive structural decomposition diagram for the entire algorithm to be solved. This graphical representation not only helps to understand the complexity of the task and the dependencies between the various parts, but also provides necessary information for subsequent resource allocation and execution sequence planning; secondly, by parsing the operator execution flowchart and determining the resource requirements of each operator unit based on a preset operator library and allocating corresponding computing resources to it, that is, each operator only requests the amount of computing resources it actually needs, and avoids using a single hardware configuration or runtime environment to process all steps, which can avoid excessive reservation and idleness of computing resources and waste of computing resources. The intelligent computing center can thereby achieve precise control and efficient use of computing resources, improve the utilization rate of computing resources, and thereby improve the computing power of the intelligent computing center, ensuring the high-performance operation of the intelligent computing center. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] 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:

[0036] Figure 1 A flowchart of a computing power scheduling method for an intelligent computing center provided by an embodiment of the present invention;

[0037] Figure 2 A flowchart of a computing power scheduling method for an intelligent computing center provided by an embodiment of the present invention;

[0038] Figure 3 A structural block diagram of a computing power scheduling device for an intelligent computing center provided by an embodiment of the present invention;

[0039] Figure 4 A structural block diagram of a computing power scheduling device for an intelligent computing center provided by an embodiment of the present invention;

[0040] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] 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.

[0042] The technical terms involved in the present invention are briefly described below.

[0043] The "computing power" mentioned in the present invention 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 requirements. It is 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. It mainly provides services to the society through computing power infrastructure.

[0044] The "computing power (CP)" mentioned in the present invention 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 per second (Floating Point Operations Per Second, FLOPS, where 1EFLOPS=10^18FLOPS). The larger the value, the stronger the comprehensive computing power. According to calculations, 1EFLOPS is approximately the computing power output of 5 Tianhe-2A or 500,000 mainstream server CPUs (Central Processing Unit) or 2 million mainstream notebooks. The calculation formula is: CP=CP 通用 +CP 智能 +CP 超级 In the embodiment of the present invention, the computing power adopts the half-precision floating point computing power (FP16) of the graphics card.

[0045] The "carrying capacity" (Network Power, NP) described in the present invention is a manifestation of the data transmission capability of computing facilities, including comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling, etc. Carrying capacity involves network transmission within and between data centers, and is a comprehensive indicator for measuring network transmission scheduling capabilities. In the embodiment of the present invention, the carrying capacity uses the video memory bandwidth. In the embodiment of the present invention, the storage capacity uses the video memory bandwidth.

[0046] The "Storage Power" (SP) described 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. In the embodiment of the present invention, the storage power adopts the video memory capacity.

[0047] The "computing power infrastructure" described in the present invention is a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity. It 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.

[0048] The “computing power” mentioned in the present invention includes general computing power CP 通用 、Intelligent computing power CP 智能 And super computing power CP 超级 .

[0049] The "general computing power" mentioned in the present invention refers to the computing power provided by servers based on CPU chips, which is used to support basic general computing such as cloud computing and edge computing.

[0050] The "intelligent computing power" mentioned in the present invention refers to the large-scale deployment of intelligent computing centers based on special chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit) for various innovative artificial intelligence applications, such as natural language processing, machine vision, etc.

[0051] The "super computing power" mentioned in the present invention is mainly the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for calculations in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0052] 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.

[0053] The "computing resources" mentioned in the present invention refer to the technologies and facilities with information calculation, transmission, storage and application capabilities required for the development of the 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.

[0054] The "operator" described in the present invention refers to a function or instruction unit that performs a specific operation, similar to a node in a flowchart. An operator is a function or instruction unit that performs a specific operation, usually working at a lower level of abstraction. An operator can be regarded as a basic component of an algorithm, usually used to perform specific computing tasks, such as mathematical operations such as addition, subtraction, multiplication, and division, or operations such as comparison and logical judgment. In programming, an operator can be a simple operator or a function that performs more complex operations (such as a database query operator or an image processing operator). Each operator unit performs an operation, such as data loading, calculation processing, etc. Operators are connected to each other through lines, which can represent the path of data flow, that is, the output of one operator unit can become the input of another operator unit. By establishing an operator execution flowchart, the structure and steps of the computing task can be clearly visualized, helping users understand the dependencies and data flow paths between the various components of the algorithm to be solved, thereby optimizing the algorithm design and subsequent resource allocation strategy.

[0055] Please refer to Figure 1 , Figure 1 A computing power scheduling method for an intelligent computing center provided according to an embodiment of the present application is shown. Figure 1 As shown, the method includes:

[0056] Step S1: Obtain an operator execution flow chart of an algorithm for a problem to be solved;

[0057] The operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line serves as the input of the operator unit corresponding to the end point of the line;

[0058] Step S2: parsing the operator execution flow chart to determine the execution order of all operator units;

[0059] Step S3: Based on the preset operator library, determine the computing power resources corresponding to each operator unit in all operator units respectively;

[0060] Step S4: allocate corresponding computing resources to each operator unit, and execute each operator unit in execution order.

[0061] In step S1, it is necessary to obtain an operator execution flowchart of the algorithm for the problem to be solved, and the operator execution flowchart describes the components of the computing task in detail.

[0062] An algorithm is a series of steps to solve a problem or perform a task. It is a process consisting of multiple operators in a specific order. In computer science, algorithms are used to describe in detail how to complete a task, such as sorting, searching, or data compression.

[0063] In a possible implementation, step S1 includes: step S11: calling a canvas tool to create a canvas; step S12: receiving an operation instruction from a user, and determining an operator execution flowchart on the canvas according to the operation instruction, wherein the operation instruction includes: an operation of selecting an operator unit from a preset operator library and an operation of connecting between operator units.

[0064] It should be noted that a canvas tool is initialized first. The canvas tool is a graphical interface component that allows users to perform graphical operations and editing. As a platform for users to interact with the system, the canvas tool provides a visually editable area in which users can create and arrange operator units. The creation of the canvas tool is the basis for users to perform intuitive operations. It provides a flexible interface that allows users to visually construct and understand the execution flow of the algorithm.

[0065] Secondly, the system responds to the user's input instructions, which include selecting operator units from the preset operator library and dragging and dropping them onto the canvas, as well as creating connections between individual operator units to define the input-output relationship between them. Each user operation will be reflected instantly on the canvas, updating the visual representation of the execution flowchart.

[0066] Therefore, by allowing users to operate operator units and connections directly on the canvas, the user's operation intuitiveness and interactivity are enhanced. Users can easily build, modify and optimize the execution process of the algorithm and see the results of the modification in real time. Such a feedback mechanism can help users better understand and design complex computing processes, and this method is suitable for application scenarios that require customization and frequent modification of computing processes, such as data science, machine learning experiments, and simulation of complex systems.

[0067] In step S2, the operator execution flowchart needs to be parsed to determine the execution order of all operator units. In other words, the operator execution flowchart obtained in step S1 is analyzed to parse out the execution order that each operator unit should follow, which usually involves sorting data dependencies to ensure that the operator that generates the data has completed execution before the operator that requires the data as input is executed. Correctly determining the execution order is the key to ensuring the correct execution of computing tasks, which can avoid errors in data processing, such as using data that has not yet been calculated, to ensure the correct solution of the problem to be solved.

[0068] Optionally, DFS (Depth-First Search) can be used to traverse the operator execution flow chart to determine the execution order of each operator unit. DFS is an algorithm for traversing or searching a tree or graph. As the name suggests, the algorithm goes deep into the branches of the graph until it cannot go any deeper, then backtracks to the nearest fork point and tries other possible paths.

[0069] In step S3, the computing resources corresponding to each operator unit in all operator units need to be determined based on the preset operator library.

[0070] Among them, the preset operator library is a predefined collection that contains a series of operator units that can be used to build and execute computing tasks, and also contains metadata information of these operator units. These operator units are usually the implementation of specific functions, which can be simple mathematical operations, complex data conversion processes, or specific algorithm implementations. The preset operator library is designed to improve development efficiency and ensure the standardization of operator implementations, so that users or intelligent computing centers can directly reference and use these operator units without having to write all the functional codes from scratch.

[0071] In step S3, it is necessary to match the required computing resources for each operator unit according to the pre-defined operator library. The operator library contains the resource information required by various operators, such as runtime, memory, GPU, CPU, etc. By accurately matching operator requirements and resources, the resource utilization efficiency can be optimized and resource waste can be reduced. At the same time, it ensures that each operator unit has sufficient computing resources for efficient calculations.

[0072] The runtime refers to the execution environment and life cycle that a software program relies on from the time it is started to the end of its execution. The runtime environment provides the program with the necessary hardware and software resources so that the program can run on the computer. This includes memory management, processor time allocation, access rights to hardware interfaces, etc.

[0073] In one possible implementation, step S3 includes: step S31: based on a preset operator library, determining the metadata information corresponding to each operator unit; step S32: based on the metadata information, determining the computing power resources corresponding to each operator unit, wherein the metadata information includes at least one of the following: the name of the operator unit, the algorithm introduction corresponding to the operator unit, the computing power resources corresponding to the operator unit, the input parameters corresponding to the operator unit, the input data set of the operator unit, the output data set of the operator unit, and the output parameters of the operator unit.

[0074] Metadata is data about data, specifically referring to detailed descriptive information about operator units, such as the name of the operator, the introduction of the relevant algorithm, the expected input and output parameters, etc. This information provides the necessary background and guidance for subsequent resource allocation and operator execution. By obtaining detailed metadata for each operator unit, the intelligent computing center can better understand the functions and requirements of each operator unit, so as to make more accurate decisions in the resource allocation and execution process. Based on the collected metadata information, each operator unit is allocated the specific computing resources it needs (which may include CPU time, memory capacity, storage space or GPU processing power, etc.). Information in the metadata, such as the computational complexity of the operator, the amount of input and output data, etc., will directly affect the decision of resource allocation. By accurately allocating resources according to the specific needs of each operator, over-allocation or under-allocation of resources can be avoided, thereby improving resource utilization efficiency and processing speed, ensuring high-performance operation of the system, and reducing computing delays and other problems caused by waste or insufficient resources.

[0075] Step S4 is the actual resource allocation and operator execution step. According to the resource requirements and execution order determined in the previous steps, the intelligent computing center will allocate corresponding resources to each operator unit and control the operators to execute in the correct order. This step is the actual execution stage of the entire computing task. Good resource allocation and strict execution order control can ensure the efficient and accurate completion of the computing task.

[0076] Please refer to Figure 2 In one possible implementation, Figure 2 As shown, step S4 includes:

[0077] Step S41: parameterize the computing power resources to obtain parameterized computing power resources;

[0078] Step S42: Based on the parameterized computing resources, call the container orchestration management tool to allocate corresponding computing resources to each operator unit, and execute each operator unit in the execution order.

[0079] Among them, step S42 includes: step S421: based on the parameterized computing power resources, build a configuration file of the container orchestration management tool; step S422: according to the configuration file and the parameterized computing power resources, create the job resources of the container orchestration management tool; step S423: call the computing power resources according to the job resources, and execute each operator unit in the execution order.

[0080] It should be noted that, first of all, the system parameterizes all possible computing resources (such as the number of CPU cores, memory size, storage space, etc.) to convert computing resource requirements into configurable and quantifiable forms so that they can be dynamically allocated and adjusted. Parameterized resources make resource configuration more flexible and controllable, facilitating rapid adjustment between different operating environments or different task requirements, improving resource utilization, and being able to accurately allocate resources according to actual needs to avoid waste of computing resources.

[0081] Secondly, based on the parameterized computing resources, call the container orchestration management tool (kubernetes) to allocate corresponding computing resources to each operator unit, and execute each operator unit in the execution order. Specifically, it is necessary to create a configuration file (such as a manifest file), which includes the resource parameters, dependencies and other running configurations required for each operator unit. This configuration file will serve as the basis for the container orchestration tool to schedule containers; then, the container orchestration management tool creates corresponding container instances for each operator unit according to the configuration file. These container instances have all the resources and configurations required to execute specific operator units; finally, after the container instances are ready, the container orchestration management tool will schedule these containers to execute according to the established execution order, that is, request and use the computing resources allocated to them. For example, if it is determined that the execution of a certain operator algorithm requires 2 nodes, each node has 8 cores and 64G, and the GPU requires a certain model, then you can create a kubernetes Job resource based on the above parameters, set the number of Job copies to 2, allocate 8 cores and 64G to it, and bind the GPU to the corresponding resources.

[0082] In summary, the generation of configuration files makes container deployment and management more standardized and automated, reduces manual configuration errors and improves deployment efficiency; by automatically creating container instances, the system can ensure that each operator unit runs in an environment tailored for it to meet its resource and configuration requirements; and ultimately ensures the execution order of computing tasks and the accuracy of resource allocation, so that each operator unit can run efficiently according to the predetermined plan, thereby optimizing the overall execution time and resource consumption.

[0083] In a possible implementation, step S4 further includes:

[0084] Step S43: after the current operator unit is executed, generate the execution status corresponding to the current operator unit, where the execution status includes any one of the following: successful execution, failed execution and executing; Step S44: determine the next operator unit to be executed according to the execution status and execution order.

[0085] It should be noted that after each operator unit is executed, the system will record its execution status. This status usually has two possibilities: successful execution, failed execution, and executing. Successful execution means that the operator unit has completed all tasks as expected without errors or exceptions. Failed execution means that an error occurred during the execution process, causing the operator unit to fail to complete its scheduled task. If the execution is successful, the system may continue to execute the next operator unit and process it in the predetermined order; if the execution fails, the system may take different measures, such as retrying the current operator unit, skipping the current operator unit, or executing some error handling logic (for example, recording errors, notifying users, etc.).

[0086] In summary, first, an operator execution flowchart is created to provide a clear and intuitive structural decomposition diagram for the entire algorithm to be solved. This graphical representation not only helps to understand the complexity of the task and the dependencies between the various parts, but also provides necessary information for subsequent resource allocation and execution sequence planning; secondly, by parsing the operator execution flowchart and determining the resource requirements of each operator unit based on the preset operator library and allocating corresponding computing resources to it, that is, each operator only requests the amount of computing resources it actually needs, and avoids using a single hardware configuration or runtime environment to process all steps. This can avoid excessive reservation and idleness of computing resources, and avoid waste of computing resources. The intelligent computing center can use this to achieve precise control and efficient use of computing resources, improve the utilization rate of computing resources, and then improve the computing power of the intelligent computing center, ensuring the high-performance operation of the intelligent computing center.

[0087] Please refer to Figure 3 , Figure 3 A structural block diagram of a computing power scheduling device for an intelligent computing center provided according to an embodiment of the present application is shown.

[0088] First, an operator library needs to be established to store operator execution code and operator metadata information. Among them, the operator is the basic algorithm execution unit for computers to solve problems. The operator library can contain the following metadata information: Operator name: such as linear binary classification algorithm operator, SFT (Supervised fine-tuning, supervised fine-tuning operator); Operator introduction: introduction to the operator algorithm; Runtime environment: the runtime required to execute the operator algorithm, such as Python Runtime, Spark Runtime; Operator input parameters: the input parameters required to execute the operator algorithm, such as the input parameters required by SFT - training rounds, learning rate, etc.; Input data set: the input data set required by the operator algorithm; Output data set: the result data set generated after executing the operator algorithm; Computing power resources: the default computing power resources required to execute the operator algorithm - CPU, memory, GPU, etc.

[0089] The front end can use the canvas form, with operators as algorithm units, and determine the order of algorithm execution between operators through connections, and the entire problem solving is transformed into a directed graph. The back end is divided into a canvas parsing module, a scheduling module, and an operator algorithm execution module. Through the cooperation of these modules, a complete problem solving process can be constructed based on the front end directed graph, and the computing power resources required to execute a single operator algorithm can be determined based on the metadata stored in the operator library, and scheduling can be performed based on the actual computing power resources used. Kubernetes can be used for algorithm scheduling, and the scheduling process and resource allocation can be determined based on the directed graph.

[0090] In other words, the problem flowchart can be solved based on the front-end canvas to determine the complete execution plan. The flowchart can be regarded as a directed acyclic graph, and the DFS algorithm can be used to traverse the nodes (operators) in the graph and determine the execution order of the operators. After the operator is executed, an output data set and output parameters will be generated. The output data set can be stored in HDFS (Hadoop Distributed FileSystem), and the flowchart status can be sent as a message to MQ (Message Queue). Then, the subsequent operators to be executed and whether the entire process is completed can be determined based on the flowchart status in MQ.

[0091] Among them, HDFS is a highly scalable distributed storage system for storing large amounts of data, and is particularly suitable for distributed computing environments. MQ is a communication method between applications that can temporarily store messages to be sent until the recipient is ready to receive them. In the computing power scheduling method of the intelligent computing center provided in the embodiment of the present application, the status of the flowchart (such as which operators have been completed and which are to be executed) can be sent to MQ as a message.

[0092] After the canvas parsing module parses the operator execution flow chart, the execution plan for the entire problem can be determined. When a specific operator algorithm step is executed, the metadata information of the operator is queried from the operator library. The following metadata will be related to the computing power allocation: the runtime of the operator (Python Runtime, Spark Runtime, etc.) to determine the environment for executing the operator algorithm; the input data set size to determine the memory to be allocated; the computing power resource parameters to determine the default CPU, GPU, and / or memory required by the operator. The computing power resources required for the final operator can be determined based on the above metadata. Parameterize the required computing power resources and dynamically build the kubernetes manifest. For example, if it is determined that the execution of a certain operator algorithm requires 2 nodes, 8 cores and 64G per node, and a certain model of GPU is required, then a kubernetes Job resource can be created based on the above parameters, the number of Job copies can be set to 2, and 8 cores and 64G are allocated to it, and the GPU is bound to the corresponding resources.

[0093] In summary, different algorithms can be encapsulated into operators (such as clustering algorithm operators, fine-tuning algorithm operators for a certain basic large model), and an operator library can be built. For a specific computer problem, it can be split into different algorithm steps, each step uses an adapted operator, and uses the appropriate runtime and allocates corresponding computing resources according to the metadata information of the operator. Decomposing computer problems into algorithm units based on operators and dynamically allocating computing resources according to the metadata information of operators can avoid the waste of computing resources.

[0094] Please refer to Figure 4 ,like Figure 4 As shown, the embodiment of the present invention further provides a computing power scheduling device for an intelligent computing center, wherein the device 40 includes:

[0095] The acquisition module 41 is used to acquire the operator execution flow chart of the algorithm to be solved; wherein the operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line is used as the input of the operator unit corresponding to the end point of the line;

[0096] The execution module 42 is used to parse the operator execution flow chart to determine the execution order of all operator units; based on the preset operator library, the computing power resources corresponding to each operator unit in all operator units are determined respectively, the corresponding computing power resources are allocated to each operator unit respectively, and each operator unit is executed in the execution order.

[0097] In a possible implementation, the acquisition module 41 is also used to call a canvas tool to create a canvas; receive a user's operation instruction, and determine an operator execution flowchart on the canvas according to the operation instruction, wherein the operation instruction includes: an operation of selecting an operator unit from a preset operator library and an operation of connecting between operator units.

[0098] In one possible implementation, the execution module 42 is also used to determine the metadata information corresponding to each operator unit based on a preset operator library; based on the metadata information, determine the computing power resources corresponding to each operator unit, wherein the metadata information includes at least one of the following: the name of the operator unit, the algorithm introduction corresponding to the operator unit, the computing power resources corresponding to the operator unit, the input parameters corresponding to the operator unit, the input data set of the operator unit, the output data set of the operator unit, and the output parameters of the operator unit.

[0099] In a possible implementation, the execution module 42 is also used to parameterize the computing power resources to obtain parameterized computing power resources; based on the parameterized computing power resources, the container orchestration management tool is called to allocate corresponding computing power resources to each operator unit, and each operator unit is executed in the execution order.

[0100] In one possible implementation, the execution module 42 is also used to build a configuration file of the container orchestration management tool based on the parameterized computing power resources; create job resources of the container orchestration management tool according to the configuration file and the parameterized computing power resources; call the computing power resources according to the job resources, and execute each operator unit in the execution order.

[0101] In a possible implementation, the execution module 42 is also used to generate an execution status corresponding to the current operator unit after the current operator unit is executed, wherein the execution status includes any one of the following: successful execution, failed execution, and executing; and determine the next operator unit to be executed based on the execution status and execution order.

[0102] In an embodiment of the present invention, an operator execution flowchart is first created to provide a clear and intuitive structural decomposition diagram for the entire algorithm to be solved. This graphical representation not only helps to understand the complexity of the task and the dependencies between the various parts, but also provides necessary information for subsequent resource allocation and execution sequence planning; secondly, by parsing the operator execution flowchart and determining the resource requirements of each operator unit based on a preset operator library and allocating corresponding computing resources to it, that is, each operator only requests the amount of computing resources it actually needs, and avoids using a single hardware configuration or runtime environment to process all steps, which can avoid excessive reservation and idleness of computing resources and waste of computing resources. The intelligent computing center can thereby achieve precise control and efficient use of computing resources, improve the utilization rate of computing resources, and thereby improve the computing power of the intelligent computing center, ensuring the high-performance operation of the intelligent computing center.

[0103] Please refer to Figure 5 An embodiment of the present invention further provides an electronic device 50, including a processor 51, a memory 52, and a computer program stored in the memory 52 and executable on the processor 51. When the computer program is executed by the processor 51, each process of the computing power scheduling method embodiment of the above-mentioned intelligent computing center is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0104] The embodiment of 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 of the above-mentioned intelligent computing center computing power scheduling method embodiment 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.

[0105] An embodiment of the present invention also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes of the computing power scheduling method embodiment of the above-mentioned intelligent computing center are implemented, and the same technical effect can be achieved. To avoid repetition, they will not be repeated here.

[0106] 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.

[0107] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods 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, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0108] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods 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 computing power scheduling method for an intelligent computing center, characterized in that: The method comprises: Step S1: obtaining an operator execution flow chart of an algorithm for a problem to be solved; wherein the operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line serves as the input of the operator unit corresponding to the end point of the line; Step S2: parsing the operator execution flow chart to determine the execution order of all operator units; Step S3: Based on the preset operator library, respectively determine the computing power resources corresponding to each operator unit in all the operator units; Step S4: Allocate corresponding computing resources to each operator unit respectively, and execute each operator unit in the execution order.

2. The method according to claim 1, characterized in that The step S1 comprises: Step S11: calling a canvas tool to create a canvas; Step S12: receiving an operation instruction from a user, and determining the operator execution flowchart on the canvas according to the operation instruction, wherein the operation instruction includes: an operation of selecting an operator unit from the preset operator library and an operation of connecting lines between the operator units.

3. The method according to claim 1, characterized in that The step S3 comprises: Step S31: Based on the preset operator library, metadata information corresponding to each operator unit is determined respectively; Step S32: Based on the metadata information, determine the computing power resources corresponding to each operator unit respectively, wherein the metadata information includes at least one of the following: the name of the operator unit, the algorithm introduction corresponding to the operator unit, the computing power resources corresponding to the operator unit, the input parameters corresponding to the operator unit, the input data set of the operator unit, the output data set of the operator unit, and the output parameters of the operator unit.

4. The method according to claim 1, characterized in that: The step S4 comprises: Step S41: parameterizing the computing power resources to obtain parameterized computing power resources; Step S42: Based on the parameterized computing resources, call the container orchestration management tool to allocate corresponding computing resources to each operator unit, and execute each operator unit in the execution order.

5. The method according to claim 4, characterized in that The step S42 comprises: Step S421: constructing a configuration file of the container orchestration management tool based on the parameterized computing resources; Step S422: Creating job resources of the container orchestration management tool according to the configuration file and the parameterized computing resources; Step S423: calling the computing power resources according to the job resources, and executing each operator unit in the execution order.

6. The method according to claim 4, characterized in that The step S4 further comprises: Step S43: after the current operator unit is executed, generating an execution status corresponding to the current operator unit, wherein the execution status includes any one of the following: successful execution, failed execution, and executing; Step S44: Determine the next operator unit to be executed according to the execution state and the execution order.

7. A computing power scheduling device for an intelligent computing center, characterized in that: The device comprises: An acquisition module is used to acquire an operator execution flow chart of an algorithm for a problem to be solved; wherein the operator execution flow chart includes a plurality of operator units and lines for connecting the operator units, and the output of the operator unit corresponding to the starting point of the line serves as the input of the operator unit corresponding to the end point of the line; An execution module is used to parse the operator execution flow chart to determine the execution order of all operator units; based on a preset operator library, the computing power resources corresponding to each operator unit in all the operator units are determined respectively, the corresponding computing power resources are allocated to each operator unit respectively, and each operator unit is executed in the execution order.

8. An electronic device, characterized in that: include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps of a method for scheduling computing power in an intelligent computing center as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a computing power scheduling method for an intelligent computing center as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that It includes computer instructions, which, when executed by a processor, implement the steps of a computing power scheduling method for an intelligent computing center as described in any one of claims 1 to 6.

Citation Information

Cited By

  • Method and device for performing asynchronous parallel fine adjustment on large model based on computing power of intelligent computing center

    CN120104277A