Affinity-first resource intelligent mapping method, equipment, medium and product

Abstract

The invention discloses an affinity-first resource intelligent mapping method, equipment, a medium and a product. The method comprises the following steps: carrying out diversity division on heterogeneous mixed computing power resources in a computing power resource pool; the method comprises the following steps: according to a computing power demand of a video large-scale scene algorithm instance, splitting the computing power demand into a series of sub-computing power demands which are connected in an associated topology manner; constructing a mapping rule with priority of computing power affinity; and sequentially and preferentially mapping the sub-computing power requirements to a logic resource pool according to the computing power affinity, and merging the computing power with different precisions by adopting a dynamic planning method to form an optimal resource combination scheme meeting the requirements of an algorithm instance. According to the affinity-first intelligent resource mapping method, computing power resource scheduling is optimized, and the computing efficiency, the resource utilization rate and the system flexibility are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power resource allocation, and particularly to a resource intelligent mapping method, device, medium and product with affinity priority. Background Art

[0002] With the development of heterogeneous computing power technology, heterogeneous computing power resources can adapt to various application scenarios. In a large-scale heterogeneous hybrid computing network, key technical challenges such as unified measurement of computing power, matching of computing power demand tasks, and dynamic scheduling of computing power resources are faced in realizing unified management and collaborative computing of homogeneous and cross-domain heterogeneous hybrid computing power resources.

[0003] To improve the utilization rate of computing power and matching accuracy, various methods have been proposed. For example, the Hybrid Metric Method (HMM) designed by Beijing Jiaotong University combines static and dynamic indicators to measure computing power resources, improving resource utilization rate and matching accuracy; Northeastern University and China Electronics Technology Group Corporation et al. integrate the advantages of NFV and SDN, and propose a general resource mapping and reuse mechanism based on SDN-NFV to achieve efficient reuse of computing power resources; the University of Athens formulates the optimal network cloud mapping problem as a Mixed Integer Programming (MIP) problem to meet the virtual resource requests with user qos awareness and effectively map computing power resource requests.

[0004] However, video large-scene algorithms usually involve various complex tasks, such as object recognition, scene understanding, real-time processing, etc. These tasks have large differences in computing power requirements. Some require high-performance GPUs, and some require low-latency dedicated hardware resources (such as FPGAs or TPUs). Most existing computing power networks do not conduct research in combination with specific large application program computing tasks and cannot meet the diverse computing power requirements of video large-scene algorithm instances. Summary of the Invention

[0005] The main purpose of this application is to provide a resource intelligent mapping method, device, medium and product with affinity priority, aiming to solve the problem of efficient allocation of existing video large-scene algorithm instances on heterogeneous hybrid computing power resources, and improve the utilization rate of computing power resources and task execution efficiency.

[0006] To achieve the above object, this application provides the following solutions:

[0007] In the first aspect, this application provides a resource intelligent mapping method with affinity priority, including the following steps:

[0008] Perform diversity partitioning on heterogeneous hybrid computing power resources in the computing power resource pool;

[0009] According to the computing power requirements of the video large-scene algorithm instance, split the computing power requirements into a series of sub-computing power requirements that are topologically interconnected;

[0010] Construct a mapping rule with computing power affinity priority, where the mapping rule includes the affinity between the computing power types required for algorithm instantiation and heterogeneous hybrid resources, as well as the computing power scale required for algorithm instantiation;

[0011] Sequentially map the sub-computing power requirements to the logical resource pool according to the computing power affinity priority. Among them, for computing powers with different precisions, a dynamic programming method is used for merging to form an optimal resource combination scheme that meets the algorithm instance requirements.

[0012] Optionally, the diversity division is based on the differentiated requirements of different data type operations for computing power resources, including but not limited to computing power types such as double-precision floating-point, single-precision floating-point, half-precision floating-point, and integer.

[0013] Optionally, the types of the sub-computing power requirements include:

[0014] Aggregate the computing power resources good at integer calculation into general computing resources mainly carried by the CPU;

[0015] Aggregate the computing power resources good at double-precision floating-point calculation into supercomputing resources mainly carried by the DCU and GPU;

[0016] Aggregate the computing power resources good at half-precision floating-point and single-precision floating-point calculation into intelligent computing resources mainly carried by the NPU.

[0017] Optionally, the construction of the mapping rule with computing power affinity priority further includes:

[0018] Adopt virtualization methods to uniformly abstract and encapsulate heterogeneous hybrid computing power resources to form an integrated computing power logical resource pool that can be uniformly scheduled and managed;

[0019] Design a computing power affinity matrix, which is used to measure the adaptability between sub-tasks and computing resources.

[0020] Optionally, the sequential mapping of the sub-computing power requirements to the logical resource pool according to the computing power affinity priority further includes:

[0021] Intelligently match the corresponding resource parameters according to the computing power affinity priority rule;

[0022] Use the greedy algorithm to sequentially select the resources with the highest computing power affinity value to execute tasks.

[0023] Optionally, the merging of computing powers with different precisions using the dynamic programming method further includes:

[0024] Define the state f(k, R′) to represent the minimum cost when the first k sub-computing power requirements are in the resource remaining state R′.

[0025] Solve according to the state transition equation to obtain the optimal resource allocation scheme, thereby determining the resources corresponding to each sub-computing power requirement.

[0026] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the resource intelligent mapping method.

[0027] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the resource intelligent mapping method.

[0028] In a fourth aspect, the present application provides a computer program product, including a computer program, wherein the computer program, when executed by a processor, implements the steps of the resource intelligent mapping method.

[0029] Through the above technical solutions, the beneficial effects of the present invention are as follows: The affinity-priority resource intelligent mapping method of the present invention optimizes the computing power resource scheduling, improves the computing efficiency, resource utilization rate, and system flexibility. The affinity-priority principle enables the computing power to preferentially select the most matching resources, avoiding the situation where some computing power devices are idle for a long time while other devices are overloaded, allowing the algorithm instances to run on the optimal hardware, and reducing the additional computing time caused by resource mismatch during the task execution process. At the same time, this allocation method promotes the development of the shared computing mode, enabling the computing power resources to better serve different applications instead of being occupied by a specific task for a long time. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0031] Figure 1 It is a flowchart of a resource intelligent mapping method with affinity priority provided by an embodiment of the present application;

[0032] Figure 2 It is a flowchart of mapping computing power requirements to heterogeneous hybrid resources based on affinity provided by an embodiment of the present application;

[0033] Figure 3 A structural schematic diagram of a computer device provided by an embodiment of the present application.

[0034] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0035] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0036] Regarding the foregoing and other technical contents, features and effects of the present invention, they will be clearly presented in the following detailed description of the embodiments in conjunction with the accompanying drawings. The structural contents mentioned in the following embodiments are all referenced to the accompanying drawings of the specification. Figures 1-3 In the following detailed description of the embodiments, it will be clearly presented. The structural contents mentioned in the following embodiments are all referenced to the accompanying drawings of the specification.

[0037] The exemplary embodiments of the present invention will be described below with reference to the accompanying drawings.

[0038] Refer to Figure 1 , an affinity - priority resource intelligent mapping method provided by an embodiment of the present application, which can be executed by a computing device with computing functions, and the computing device can be a desktop computer, a laptop computer, etc.

[0039] Step S1: Perform diversity partitioning on the heterogeneous mixed computing power resources in the computing power resource pool;

[0040] Step S2: According to the computing power requirements of the video large - scene algorithm instance, split the computing power requirements into a series of sub - computing power requirements with associated topological interconnections;

[0041] Step S3: Construct a mapping rule with priority for computing power affinity, where the mapping rule includes the affinity between the computing power types required for algorithm instantiation and the heterogeneous mixed resources, and the computing power scale required for algorithm instantiation;

[0042] Step S4: Sequentially map the sub - computing power requirements to the logical resource pool with priority for computing power affinity. Among them, dynamic programming is used to merge the computing power of different precisions to form an optimal resource combination scheme that meets the requirements of the algorithm instance.

[0043] In step S1, the diversity partitioning is based on the different requirements of computing power resources for operations of different data types, including but not limited to computing power types such as double - precision floating - point, single - precision floating - point, half - precision floating - point, and integer.

[0044] The computing precision of the above computing types is usually determined by the following factors: numerical precision, algorithm complexity, and computational error tolerance. Algorithms with different precision requirements will affect the computing power demand, resulting in differences in the allocation of computing resources.

[0045] In step S2, for a given video large-scenario algorithm instance, it is split according to different computing types. Data preprocessing (integer computing) is allocated to CPU resources; feature extraction (vector computing) is allocated to GPU resources; deep learning inference (tensor computing) is allocated to NPU resources; post-processing (logical judgment) is allocated to CPU resources. The mathematical model is as follows:

[0046] T = {T1, T2,..., T n}

[0047] where T represents the total computing power demand of the entire algorithm instance, and T i represents the computing power demand of a certain subtask. Each subtask has different computing types, such as integer computing, floating-point computing, vector computing, and tensor computing.

[0048] The algorithm instance is split into multiple related subtasks and a computing topology graph G is formed:

[0049] G = (V, E)

[0050] where V represents each subtask T i , that is, the computing power demand, and E represents the data dependency relationship between tasks, that is, the order of task execution.

[0051] In a specific implementation, the types of subtask computing power demands include:

[0052] Aggregate the computing power resources good at integer computing into general computing resources mainly carried by the CPU;

[0053] Aggregate the computing power resources good at double-precision floating-point computing into supercomputing resources mainly carried by the DCU and GPU;

[0054] Aggregate the computing power resources good at half-precision floating-point and single-precision floating-point computing into intelligent computing resources mainly carried by the NPU.

[0055] Table 1 Principles for Dividing Computing Power Resources

[0056]

[0057] In the above table, to map different algorithm instances to efficient computing resources, the computing power resources in the resource logical pool are diversely divided. Based on the adaptation capabilities of different hardware for different types of computing tasks, the computing power resources can be divided into: general computing resources (dominated by CPUs), supercomputing resources (dominated by DCUs / GPUs), and intelligent computing resources (dominated by NPUs). In computing resource management, the resource logical pool aggregates various computing resources to provide flexible computing power allocation for different computing tasks. Since the types, computing requirements, and resource characteristics of computing tasks are different, the present invention diversely divides the computing power resources to achieve efficient computing, optimize resource utilization, and enhance the computing power of the overall system.

[0058] As Figure 2 shown, in the exemplary embodiment, it is assumed that there is a video processing algorithm instance, and this scenario requires real-time object detection, behavior analysis, and intelligent classification of video content to achieve rapid early warning of abnormal behaviors and efficient management of video data. The computing power requirements of the entire video surveillance system are split into a series of associated topologically interconnected sub-computing power requirements according to different functional modules. Specifically, the acquisition and decoding part of the video stream is defined as sub-task T1, which mainly involves integer computing and is responsible for by the CPU; the object detection part is defined as sub-task T2, which involves double-precision floating-point computing and vector processing and is processed by the DCU or GPU; the behavior analysis part is defined as sub-task T3, which involves half-precision floating-point and single-precision floating-point computing and is processed by the NPU; the intelligent classification part of the video content is defined as sub-task T4, which also involves half-precision floating-point and single-precision floating-point computing and is also processed by the NPU. There are data dependency relationships between these sub-tasks. For example, only after the video decoding (T1) is completed can the object detection (T2) be performed, and the result of the object detection will be used as the input data for the behavior analysis (T3) and the intelligent classification of video content (T4).

[0059] In step S3, further including constructing a mapping rule with computing power affinity priority:

[0060] Using virtualization methods to uniformly abstract and encapsulate heterogeneous hybrid computing power resources, shielding the differences in hardware architectures and instruction sets of computing power resources, and forming an integrated computing power logical resource pool that can be uniformly scheduled and managed;

[0061] Designing a computing power affinity matrix, where the computing power affinity matrix is used to measure the adaptability of sub-tasks to computing resources. Among them, the computing power affinity is calculated based on the execution performance and cost of tasks on resources.

[0062] Furthermore, further including mapping the sub-computing power requirements to the logical resource pool in sequence according to the computing power affinity priority:

[0063] Intelligently matching corresponding resource parameters according to the computing power affinity priority rule;

[0064] Use the greedy algorithm to sequentially select the resources with the highest computing power affinity value to execute tasks.

[0065] In a specific implementation, the computing power affinity is defined as follows:

[0066]

[0067] where A(T i , R j ) represents the affinity value of task T i on resource R j ; P(T i , R j ) represents the execution performance of task T i on resource R j (such as throughput, parallelism, etc.); C(T i , R j ) represents the computing cost of task T i on resource R j (such as latency, power consumption, occupancy, etc.).

[0068] By calculating the affinity values between different tasks and resources, a computing power affinity matrix is formed:

[0069]

[0070] According to the computing power affinity matrix, use the greedy algorithm to sequentially select the optimal computing resources:

[0071] R opt (T i ) = arg max A(T i , R j )

[0072] That is: Select the resource R j with the highest computing power affinity value to execute task T i .

[0073] In step S4, the further merging of computing powers with different precisions using the dynamic programming method further includes:

[0074] Define the state f(k, R′) to represent the minimum cost when the first k sub-computing power requirements are in the resource remaining state of R′;

[0075] Solve according to the state transition equation to obtain the optimal resource allocation scheme, thereby determining the resources corresponding to each sub-computing power requirement.

[0076] Specifically, let x ij be a binary variable. If the sub-computing power requirement T i is mapped to resource Rj If it is greater than, then x ij = 1, otherwise x ij = 0. The goal is to find a set of values such that the resource combination is optimal, that is, while meeting the requirements of the algorithm instance, minimizing the usage cost of resources. Let the usage cost of resource R j be cost j , then the objective function can be expressed as:

[0077]

[0078] Each sub-computing power requirement must be mapped to a certain resource:

[0079]

[0080] The allocation of resources cannot exceed its computing power supply capacity:

[0081]

[0082] Use the dynamic programming algorithm to solve the above optimization problem. Define the state f(k, R′) to represent the minimum cost when the first k sub-computing power requirements are in the resource remaining state R′. The state transition equation is as follows:

[0083]

[0084] Among them, x kj represents mapping the k-th sub-computing power requirement to resource R j .

[0085] Through dynamic programming, the optimal x ij values are obtained. According to these values, the resources corresponding to each sub-computing power requirement are determined, thereby forming an optimal resource combination plan that meets the requirements of the algorithm instance, so as to maximize the reduction of resource usage cost, improve resource utilization rate and task execution efficiency while ensuring the normal operation of the system function.

[0086] In summary, the affinity-priority resource intelligent mapping method of the present invention optimizes the computing power resource scheduling, improves the computing efficiency, resource utilization rate and system flexibility. The affinity-priority principle enables the computing power to preferentially select the most matching resources, avoiding the situation where some computing power devices are idle for a long time while other devices are overloaded, allowing the algorithm instance to run on the optimal hardware, and reducing the additional computing time caused by resource mismatch during the task execution process. At the same time, this allocation method promotes the development of the shared computing mode, enabling the computing power resources to better serve different applications instead of being occupied by a specific task for a long time.

[0087] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as shown in Figure 3 Shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store resource intelligent mapping data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a resource intelligent mapping method with affinity priority.

[0088] Those skilled in the art can understand that Figure 3 The structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0089] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0090] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0091] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0092] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0093] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0094] The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0095] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0096] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A resource intelligent mapping method with affinity priority, characterized in that: The steps include: Diversify the heterogeneous mixed computing resources in the computing resource pool; According to the computing power requirements of the video large scene algorithm instance, the computing power requirements are split into a series of sub-computing power requirements that are interconnected with each other in an associated topology; Construct a mapping rule that prioritizes computing power affinity; The sub-computing power requirements are mapped to the logical resource pool in order according to the computing power affinity, wherein the computing power of different precisions is merged using a dynamic programming method to form an optimal resource combination solution that meets the requirements of the algorithm instance.

2. The affinity-prioritized resource intelligent mapping method according to claim 1, characterized in that: The diversity division is based on the differentiated demands for computing resources for operations of different data types, including but not limited to double-precision floating point, single-precision floating point, half-precision floating point, and integer computing types.

3. The affinity-first resource intelligent mapping method according to claim 2, characterized in that: The types of sub-computing power requirements include: Aggregate computing resources that are good at integer computing into general computing resources that are mainly CPU-based; Aggregate computing resources that excel at double-precision floating-point calculations into supercomputing resources that are mainly carried by DCU and GPU; The computing resources that are good at half-precision floating-point and single-precision floating-point calculations are aggregated into intelligent computing resources mainly carried by NPU.

4. The affinity-first resource intelligent mapping method according to claim 3 is characterized in that: The mapping rule of constructing the computing power affinity priority further includes: Use virtualization methods to uniformly abstract and encapsulate heterogeneous hybrid computing resources, forming an integrated computing logic resource pool that can be uniformly scheduled and managed; A computing power affinity matrix is ​​designed, where the computing power affinity matrix is ​​used to measure the compatibility of subtasks with computing resources.

5. The affinity-first resource intelligent mapping method according to claim 4 is characterized in that: Mapping the sub-computing power requirements to the logical resource pool in order according to the computing power affinity priority further includes: Intelligently match corresponding resource parameters according to the computing power affinity priority rule; Use a greedy algorithm to select the resource with the highest computing power affinity value to execute the task.

6. The affinity-prioritized resource intelligent mapping method according to claim 5, characterized in that: The merging of computing power of different precisions by using a dynamic programming method further includes: Define the state f(k,R ′ ) indicates that the first k sub-computing power requirements are in the resource remaining state of R ′ The minimum cost when The optimal resource allocation plan is obtained by solving the state transfer equation, thereby determining the resources corresponding to each sub-computing power requirement.

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the resource intelligent mapping method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the resource intelligent mapping method described in any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the resource intelligent mapping method described in any one of claims 1 to 6 are implemented.

Images