Dynamic extensible heterogeneous collaborative intelligent computing platform and method for edge-end equipment

Abstract

The invention relates to a dynamic extensible heterogeneous collaborative intelligent computing platform and method for edge-end equipment, and relates to the technical field of edge intelligent computing and computer hardware, and the platform comprises a hardware infrastructure layer and a master control module for realizing signal transmission and resource sharing; the basic software layer is used for managing hardware resources and software services; the application support layer is used for running computing resources and realizing preliminary compatibility; the task management and resource scheduling layer is responsible for task arrangement and management and task scheduling decision; the application layer is used for providing various application services; the data transmission layer is responsible for transmitting data among different platforms; the computing resource interface layer provides an interface interacting with heterogeneous computing resources; and the computing resource platform is composed of different heterogeneous computing platforms and provides computing resources and computing power. The method has the capabilities of intelligent information and signal processing, diversification of application scenes, efficient parallel operation, rapid iterative development, universal flexible deployment and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge intelligent computing and computer hardware, in particular, to a dynamic scalable heterogeneous collaborative intelligent computing platform for edge equipment and a method thereof. BACKGROUND

[0002] With the continuous progress of information and signal processing, intelligent and other related technologies, and the increasing diversity and complexity of application scenarios, the demand for efficient parallel operation, rapid iterative development, general flexible deployment and the like of edge equipment is growing. Under such a development trend, the traditional homogeneous computing platform gradually exposes limitations that cannot meet the actual needs.

[0003] For example, the prior art discloses a high-performance heterogeneous intelligent computing platform, which constructs a multi-core heterogeneous platform by heterogeneously fusing CPU, NPU, DSP and FPGA four types of core chips, forms a general computing resource center, realizes unified deployment of various computing resources, and is suitable for full-process control of unmanned aerial vehicles in the task process. However, this technical solution mainly focuses on the heterogeneous fusion of underlying hardware facilities, which improves the computing capability to a certain extent, but still cannot effectively solve the problems of diverse and complex application scenarios, large amount of operation, slow iterative development and low scalability.

[0004] In addition, the prior art further discloses a collaborative processing system of a heterogeneous intelligent computing platform, which realizes the deployment of multiple types of intelligent tasks and the subsequent expansion and iteration of intelligent tasks by setting a general processing module, an image type computing module and a natural language processing module. Although this solution realizes the diversification and configurability of the intelligent technology platform system function to a certain extent, it still has limitations when facing the demand for complex multi-computing resource collaboration and efficient resource deployment.

[0005] The existing heterogeneous computing platform technology still has deficiencies in meeting the diverse application scenarios of edge equipment, efficient parallel operation, rapid iterative development and flexible deployment and the like. Therefore, there is an urgent need for a dynamic scalable heterogeneous collaborative intelligent computing platform that can better adapt to the needs of edge equipment to overcome the limitations of existing technologies and provide more efficient, flexible and scalable computing solutions. SUMMARY

[0006] The purpose of the present application is to provide a dynamic scalable heterogeneous collaborative intelligent computing platform and method for edge equipment, which has the capabilities of intelligent information and signal processing, diversified application scenarios, efficient parallel operation, rapid iterative development, general flexible deployment and the like.

[0007] In order to achieve the above-mentioned purpose, the technical solution of the present application provides a dynamic scalable heterogeneous collaborative intelligent computing platform for edge equipment, which comprises:

[0008] Hardware infrastructure layer: including master module, which realizes signal transmission and resource sharing;

[0009] Basic software layer: running on the hardware infrastructure layer, used to manage hardware resources and software services, provide isolated environment for application running, and ensure efficient data transmission between different modules;

[0010] Application support layer: built on the basic software layer, used to run various computing resources and implement preliminary compatibility with multiple ecological libraries;

[0011] Task management and resource scheduling layer: responsible for task arrangement and management and task scheduling decision;

[0012] Application layer: provides information and signal processing applications, intelligent application services, and meets the diversified needs of users in different scenarios;

[0013] Data transmission layer: responsible for transmitting data between different platforms, using data transmission protocol and data compression;

[0014] Computing resource interface layer: provides interfaces for interacting with heterogeneous computing resources, implements API interface standardization and communication protocol conversion and adaptation, and monitors the status of heterogeneous computing resources;

[0015] Computing resource platform: composed of different heterogeneous computing platforms, providing computing resources and computing power to execute assigned computing tasks.

[0016] Preferably, the hardware infrastructure layer adopts an open architecture, and each bottom hardware module follows a unified interface specification for interconnection to ensure the cooperative work of the entire hardware system.

[0017] Preferably, the basic software layer provides a stable and reliable running environment for the upper application support layer through close cooperation with the hardware infrastructure layer.

[0018] Preferably, the application support layer provides rich computing resources and support libraries for the task management and resource scheduling layer through cooperation with the basic software layer, ensuring efficient task execution.

[0019] Preferably, the task management and resource scheduling layer formulates corresponding scheduling strategies such as load balancing, cost optimization and real-time guarantee according to the resource status of each computing platform, i.e. CPU, GPGPU, memory and storage, and dynamically adjusts task allocation according to actual conditions.

[0020] Preferably, the data transmission layer uses data transmission protocol and data compression technology to improve data transmission speed and efficiency, while enhancing data security and encryption processing to ensure the security of the data transmission process.

[0021] Preferably, the computing resource interface layer implements the standardization of the API interface and the conversion and adaptation of the communication protocol, so that different types of computing resources such as CPU, GPGPU, NPU can work collaboratively under a unified framework.

[0022] The technical scheme of the present application also provides a dynamic scalable heterogeneous collaborative intelligent computing method for edge-end equipment, which comprises:

[0023] Priority task modeling based on directed acyclic graph, according to application task analysis subtask composition, dependency relationship and priority and construction of directed acyclic graph;

[0024] Designing an environment interaction framework to realize real-time collection and sharing of key state data such as node computing capability, network topology and transmission capability, and load between distributed nodes;

[0025] Designing a reinforcement learning model, designing simulation data and environment pre-training to build a learning model according to typical tasks such as pulsed Doppler radar and target recognition, and continuously optimizing to realize the ability to search for the global optimal scheduling strategy according to the hypothesis condition and transient state.

[0026] Preferably, the priority task modeling based on directed acyclic graph comprises: analyzing the subtask composition, dependency relationship and priority of the application task, and constructing a directed acyclic graph to clearly show the execution order and dependency relationship between tasks, providing a basis for the resource scheduling layer to formulate scheduling strategies.

[0027] Preferably, the reinforcement learning model automatically learns and adjusts the scheduling strategy according to historical scheduling data and current system state, to adapt to different task requirements and system environment.

[0028] Due to the adoption of the above technical scheme, the present application has the following advantages:

[0029] The architecture design of the present application contains multiple levels from hardware infrastructure to application layer, with the capabilities of intelligent information and signal processing, diversification of application scenarios, efficient parallel operation, rapid iterative development, general flexible deployment, etc. This design fully considers the overall needs of edge-end equipment in terms of power consumption, security, reliability, versatility, scalability and real-time performance, and through the integration of diversified computing resources, it can flexibly cope with the diversified needs of complex computing tasks for resources, effectively improving the intelligent level and task processing capability of edge-end equipment.

[0030] The dynamic expandable heterogeneous collaborative intelligent computing platform architecture of the application successfully integrates various computing resources, including CPU, GPGPU, NPU, etc. This integration method can meet the demand of complex computing tasks for computing resources through more flexible computing methods, provides a unified interface design, allows dynamic adjustment and optimization of the allocation of computing resources according to the characteristics of computing tasks and computing environment, thereby improving the utilization efficiency of computing resources and the flexibility of task execution, and enhancing the adaptability and expansibility of the platform.

[0031] The heterogeneous resource intelligent collaborative computing framework interface specification of the application preliminarily compatible with various ecological libraries, forms a set of unified support library, and can efficiently run various computing resources. The interface specification effectively solves the problem of non-uniformity of heterogeneous computing platform ecology, non-compatibility of resource calling and data transmission protocol, provides a standardized interface and communication protocol for the collaborative work of heterogeneous computing resources, reduces the complexity of system integration, and improves the maintainability and scalability of the system.

[0032] The heterogeneous resource intelligent collaborative computing framework designed by the application collects and shares the real-time collection and sharing of key state data such as node computing capability, network topology and transmission capability, and load, to provide accurate data support for task scheduling decision-making. This design enables the heterogeneous resource intelligent collaborative computing framework to make more accurate scheduling decisions according to real-time state data, improves the efficiency of task scheduling and the overall performance of the system, and enhances the applicability and value of the heterogeneous computing platform in actual application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A dynamic expandable heterogeneous collaborative intelligent computing platform for edge-end equipment of the application is shown in the figure;

[0034] Figure 2 A dynamic expandable heterogeneous collaborative intelligent computing platform architecture of the application is shown in the figure;

[0035] Figure 3 A heterogeneous resource intelligent collaborative computing framework interface specification of the application is shown in the figure;

[0036] Figure 4 A heterogeneous resource intelligent collaborative computing framework of the application is shown in the figure. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0038] The embodiment of the application discloses a dynamic scalable heterogeneous collaborative intelligent computing platform for edge equipment, which has a multi-level architecture from hardware infrastructure to application layer to realize intelligent information and signal processing, diversified application scenarios, efficient parallel operation, rapid iterative development and flexible deployment and other key capabilities. This architecture fully considers the overall needs of edge equipment in terms of power consumption, security, reliability, versatility, scalability and real-time performance, and can flexibly cope with the diversified needs of complex computing tasks for resources by integrating various computing resources such as CPU, GPGPU and NPU.

[0039] As shown in the accompanying drawings Figure 1 The computing platform architecture mainly consists of interface specifications, hardware infrastructure layer, basic software layer, application support layer, task management and resource scheduling layer and application layer. The hardware infrastructure layer constitutes the physical basis of the entire platform, which is composed of bottom hardware modules such as main control module, communication module, heterogeneous computing resources, power module and expansion module and their interconnection modes, providing necessary hardware resources for upper layer software. The basic software layer runs on the hardware infrastructure, including operating system, container, communication middleware, firmware, driver, BSP and other components, responsible for allocating and scheduling basic hardware resources for upper layer software, ensuring that the software system can efficiently utilize hardware facilities.

[0040] The application support layer is built on the basic software layer and contains intelligent computing framework and unified support library and other key components, providing necessary support environment for the task management and resource scheduling layer. The intelligent computing framework can efficiently run various computing resources, while the unified support library realizes preliminary compatibility for various ecological libraries through interface specifications, effectively solving the problem of non-uniformity of heterogeneous computing platform ecology, non-compatibility of resource calling and data transmission protocol. The application support layer provides rich computing resources and support libraries for the task management and resource scheduling layer through cooperation with the basic software layer, ensuring efficient execution of tasks.

[0041] The task management and resource scheduling layer is one of the core parts of the computing platform, responsible for task arrangement and management and task scheduling decision. This layer can parse the tasks submitted by users, split them into executable small tasks, and manage the task queue to ensure that the tasks are executed in order according to priority and dependency. At the same time, it can also evaluate the resource status of each computing platform, including CPU, GPGPU, memory, storage and other resources, formulate scheduling strategies such as load balancing, cost optimization and real-time performance guarantee, and dynamically adjust task allocation according to actual conditions to ensure orderly execution of tasks.

[0042] The application layer provides information and signal processing applications, intelligent applications, and various application services to meet the diverse needs of users in different scenarios. The user interface layer provides a friendly interface or API interface for users to submit computing tasks, query task status, and receive feedback, thereby realizing user interaction with the computing platform. The task management layer is responsible for parsing user-submitted tasks, splitting them into executable small tasks, and managing task queues to ensure that tasks are executed in order according to priority and dependency. The resource scheduling layer assesses the resource status of each computing platform, formulates scheduling strategies, executes scheduling decisions, allocates tasks to the most suitable computing platform, and adjusts according to actual conditions, such as when the load of a certain computing platform is too high, the resource scheduling layer will transfer part of the tasks to other computing platforms with lower load to achieve load balancing.

[0043] The data transmission layer is responsible for transmitting data between different platforms, using efficient data transmission protocols and data compression techniques to improve data transmission speed and efficiency, while enhancing data security and encryption processing to ensure the security of the data transmission process. The computing resource interface layer provides an interface for interacting with heterogeneous computing resources, standardizes API interfaces, and converts and adapts communication protocols, and monitors the status of heterogeneous computing resources to provide feedback to the resource scheduling layer for dynamic adjustment. The computing resource platform is composed of different heterogeneous computing platforms that provide necessary computing resources and computing capabilities to execute assigned computing tasks.

[0044] As shown in Figure 2 , the dynamic scalable heterogeneous collaborative intelligent computing platform architecture adopts an open system architecture design, and the software adopts a hierarchical and component-based layered architecture, allowing each component to develop independently. By uniformly defining the functions and external interfaces of each layer of software components, the upgrade of each layer of software components can be integrated into the system as long as the consistency of the external interface is maintained, without affecting other software components. Even if the interface of a certain layer of components needs to be changed, it can be adjusted through the corresponding interconnection interface to achieve mutual independence with other software components.

[0045] As shown in Figure 3 , the heterogeneous resource intelligent collaborative computing framework interface specification includes a user interface layer, a task management layer, a resource scheduling layer, a data transmission layer, and a computing resource interface layer. The user interface layer provides convenience for user operations, the task management layer is responsible for task parsing and queue management, the resource scheduling layer focuses on resource status assessment and scheduling strategy formulation, the data transmission layer ensures efficient and secure data transmission, and the computing resource interface layer realizes interaction with and state monitoring of heterogeneous computing resources.

[0046] Figure 4The isomer resource intelligent collaborative computing framework shown adopts a priority task modeling technology based on a directed acyclic graph, can analyze subtask composition, dependency relationship and priority according to application task and build a directed acyclic graph, wherein the accepted task is divided into ①, ②, ③, ④ and ⑤ multiple subtasks according to priority. At the same time, an environment interaction framework is designed to realize real-time collection and sharing of key state data such as node computing capability, network topology and transmission capability and load between distributed nodes, which enables the resource scheduling layer to obtain the running state of each computing platform in time, so as to make more accurate scheduling decisions. For example, when the load of a certain node is too high or the network transmission capability is limited, the resource scheduling layer can transfer the task to other more suitable nodes to improve the task execution efficiency. In addition, a reinforcement learning model is also designed, which designs simulation data and environment pre-training to build a learning model according to typical tasks such as pulsed Doppler radar and target recognition and continuously optimizes it. By abstracting the modeling search space and adaptive function elements, the ability to search for the global optimal scheduling strategy according to the hypothetical conditions and transient state is realized. The reinforcement learning model can automatically learn and adjust the scheduling strategy according to the historical scheduling data and the current system state to adapt to different task requirements and system environments.

[0047] In the hardware infrastructure layer, the master module serves as the core control unit and realizes signal transmission and resource sharing with other communication modules, heterogeneous computing resource modules, power modules and expansion modules through specific connection methods. The connection methods and signal transmission methods between these modules need to follow unified interface specifications to ensure the collaborative work of the entire hardware system. The interconnection of the underlying hardware modules constitutes the hardware infrastructure and provides necessary physical support for the upper layer of basic software.

[0048] The basic software layer runs on the hardware infrastructure, including components such as operating system, container, communication middleware, firmware, driver and BSP. The operating system is responsible for managing hardware resources and software services, the container technology provides an isolated environment for application running, the communication middleware ensures efficient data transmission between different modules, the firmware and driver ensure the normal operation of hardware devices, and the BSP (board support package) provides support for the adaptation between the operating system and the hardware platform. The basic software layer provides a stable and reliable running environment for the upper layer of application support layer through close cooperation with the hardware infrastructure layer.

[0049] In actual operation, when a user submits a computing task through the user interface layer, the task management layer parses the task and splits it into multiple executable small tasks, while managing the task queue to ensure that tasks are executed in order according to priority and dependency. The resource scheduling layer assesses the available resources, such as CPU, GPGPU, memory, storage, etc., and formulates appropriate scheduling strategies, such as load balancing, cost optimization, and real-time performance guarantee, to ensure efficient execution of tasks.

[0050] The computing resource interface layer provides an interactive interface for heterogeneous computing resources, standardizes API interfaces, and converts and adapts communication protocols. This allows different types of computing resources, such as CPU, GPGPU, NPU, etc., to work together under a unified framework. The computing resource interface layer also monitors the status of heterogeneous computing resources and provides feedback to the resource scheduling layer for dynamic adjustment. For example, when a computing resource fails or performance decreases, the computing resource interface layer will promptly notify the resource scheduling layer to reallocate tasks and ensure normal execution of tasks.

[0051] The computing resource platform, as an entity for executing allocated computing tasks, provides necessary computing resources and computing capabilities through different heterogeneous computing platforms. These platforms include CPU, GPGPU, NPU, and other computing resources, which can meet the computing needs of different types of tasks. For example, CPU can handle general computing tasks, GPGPU is suitable for parallel computing-intensive tasks, and NPU excels in artificial intelligence tasks such as deep learning.

[0052] To achieve efficient scheduling of tasks and rational utilization of resources, the invention designs a priority task modeling technology based on directed acyclic graphs. According to the analysis of subtask composition, dependency relationship, and priority of application tasks, a directed acyclic graph is constructed. In this way, the execution order and dependency relationship between tasks can be clearly displayed, providing a basis for the resource scheduling layer to formulate scheduling strategies.

[0053] The dynamic scalable heterogeneous collaborative intelligent computing platform of the embodiment of the invention achieves intelligent information and signal processing, diversified application scenarios, efficient parallel operation, rapid iterative development, and general flexible deployment through the collaborative work of the multi-layer architecture. The platform fully considers the overall needs of edge devices, integrates various computing resources and designs unified interface specifications, solves the problems of non-uniformity of heterogeneous computing platform ecology, non-compatibility of resource calling and data transmission protocols, improves the accuracy of task scheduling decisions and the response speed of the system to state changes, and has important practical application value.

[0054] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that the technical solutions described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalent features, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices, characterized in that, include: Hardware infrastructure layer: This includes the main control module, which enables signal transmission and resource sharing; Basic software layer: Running on top of the hardware infrastructure layer, it is used to manage hardware resources and software services, provide an isolated environment for application execution, and ensure efficient data transmission between different modules; Application support layer: Built on top of the basic software layer, it is used to run various computing resources and achieve initial compatibility with various ecosystem libraries; Task Management and Resource Scheduling Layer: Responsible for task orchestration and management, as well as task scheduling decisions; Application layer: Provides a variety of application services, including information and signal processing applications and intelligent applications, to meet the diverse needs of users in different scenarios; Data transmission layer: responsible for transmitting data between different platforms, using data transmission protocols and data compression; Computing Resource Interface Layer: Provides interfaces for interacting with heterogeneous computing resources, standardizes API interfaces and converts and adapts communication protocols, and monitors the status of heterogeneous computing resources; Computing resource platform: Composed of different heterogeneous computing platforms, providing computing resources and computing capabilities to execute assigned computing tasks.

2. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 1, characterized in that, The hardware infrastructure layer adopts an open architecture, and the underlying hardware modules are interconnected in accordance with a unified interface specification to ensure the collaborative operation of the entire hardware system.

3. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 2, characterized in that, The basic software layer, through close collaboration with the hardware infrastructure layer, provides a stable and reliable operating environment for the upper application support layer.

4. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 3, characterized in that, The application support layer collaborates with the basic software layer to provide the task management and resource scheduling layer with abundant computing resources and support libraries, ensuring the efficient execution of tasks.

5. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 4, characterized in that, The task management and resource scheduling layer formulates corresponding scheduling strategies such as load balancing, cost optimization, and real-time guarantee based on the resource status of each computing platform, namely CPU, GPGPU, memory, and storage, and dynamically adjusts task allocation according to the actual situation.

6. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 5, characterized in that, The data transmission layer employs data transmission protocols and data compression technologies to improve data transmission speed and efficiency, while strengthening data security and encryption to ensure security during data transmission.

7. The dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices according to claim 6, characterized in that, The computing resource interface layer standardizes API interfaces and converts and adapts communication protocols, enabling different types of computing resources, such as CPUs, GPGPUs, and NPUs, to work together under a unified framework.

8. The method of the dynamically scalable heterogeneous collaborative intelligent computing platform for edge devices as described in claim 7, characterized in that, include: Priority task modeling based on directed acyclic graphs involves analyzing the composition, dependencies, and priorities of subtasks according to the application task and constructing a directed acyclic graph. Design an environment interaction framework to realize the real-time collection and sharing of key status data such as node computing power, network topology and transmission capacity, and load among distributed nodes; We designed a reinforcement learning model, and based on typical tasks such as pulse Doppler radar and target recognition, we designed simulation data and environmental pre-training to build a learning model and continuously optimized it, so as to achieve the ability to search for the globally optimal scheduling strategy based on assumptions and transients.

9. The dynamically scalable heterogeneous collaborative intelligent computing method for edge devices according to claim 8, characterized in that, Priority task modeling based on directed acyclic graphs includes: analyzing the subtask composition, dependencies, and priorities of application tasks, and constructing a directed acyclic graph to clearly show the execution order and dependencies between tasks, providing a basis for the resource scheduling layer to formulate scheduling strategies.

10. The dynamic scalable heterogeneous collaborative intelligent computing method according to claim 9, characterized in that, The reinforcement learning model automatically learns and adjusts scheduling strategies based on historical scheduling data and the current system state to adapt to different task requirements and system environments.

Images