Edge-end-side heterogeneous computing power resource scheduling system and method

Abstract

The invention discloses an edge-end-side heterogeneous computing power resource scheduling system and method.The system comprises a computing power virtualization resource pool module, a cross-domain distributed computing power soft bus and a computing power resource scheduling management module, the computing power virtualization resource pool module is based on the heterogeneous computing power virtual computing resource pool technology, and the cross-domain distributed computing power soft bus is used for scheduling the computing power resource scheduling management module; performing unified management and virtualized mapping on various equipment systems on the edge side and the end side according to computing power resources, network resources and storage resources to form a unified virtual computing resource pool; according to the cross-domain distributed computing power soft bus, intermediate routing forwarding nodes of computing power resources are established among different communication domains, the routing forwarding nodes support a power system physical layer protocol, and computing power resource lists among different communication protocol networks are synchronously updated in real time; and the computing power resource scheduling management module is used for realizing dynamic allocation management of computing power resources based on a dynamic computing power scheduling system established by the computing power resource pool, evaluating the computing power demand required by each computing task in real time, and requesting the virtual computing resource pool to allocate an optimal computing power resource combination for algorithm application.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and specifically relates to a system and method for scheduling heterogeneous computing resources on an edge side. Background Art

[0002] At present, the new power system is developing rapidly. After household photovoltaics, charging piles, industrial and commercial energy storage, smart air conditioning and lighting and other end-side equipment are connected to the power grid, they are faced with massive data processing and analysis, real-time calculation, and concurrent control requirements. The power grid data algorithm analysis capabilities and real-time processing requirements are improved, which puts great pressure on the two-way interaction, safe and stable operation of the power grid.

[0003] At present, there are a large number of IoT access devices in the power industry. There are many types of computing resources used in data applications, and they are all used locally. It is impossible to achieve cross-system and cross-device computing migration and reuse, and the computing resource utilization rate is not high. A standardized management device and method is urgently needed to realize cross-domain coordination, scheduling, and collaborative intelligent processing applications of heterogeneous computing resources on the edge side. Summary of the invention

[0004] The purpose of the invention is to provide a system and method for scheduling heterogeneous computing power resources on the edge side in view of the above shortcomings, which is applied to the power Internet of Things and can realize the unified management of virtualization of heterogeneous computing power devices on the edge side, cross-domain distributed computing power soft bus, and dynamic scheduling management of computing power. It supports heterogeneous asymmetric computing power architecture, supports cross-domain calling and scheduling coordination of computing power resources, reduces the management cost of heterogeneous computing power resources on the edge side, and improves the utilization efficiency of computing power resources on the edge side.

[0005] The computing power resources in this system refer to various processor resources that can realize algorithm computing functions on the edge side, including graphics processing units GPU, video processing units VPU, neural network processors NPU and central processing units CPU, etc. The computing power index parameters can be used to measure the computing power of different computing power resources.

[0006] The present invention is implemented by the following technical solutions: Heterogeneous computing power virtual computing resource pool management includes heterogeneous computing power virtual computing resource pool technology, which can uniformly manage and virtualize various device systems on the edge side according to computing power resources, network resources and storage resources to form a unified virtual computing resource pool. Different types and manufacturers of edge devices are abstracted into different virtual computing resources according to their processing capabilities, and computing power resources (GPU, NPU, TPU and VPU), network resources (Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485), storage resources (physical storage, virtual storage, shared storage) are configured on different virtual computing resources according to the physical connection method, and their computing power indicators (number of floating-point operations per second, number of instructions executed per second and video memory parameters) are quantified to carry different computing power tasks. Through the elastic management of the virtual computing resource pool, it supports dynamic configuration, elastic expansion, load balancing, fault detection and other mechanisms of computing power resources to ensure the high availability and operation reliability of computing power resources during the execution of computing tasks.

[0007] The virtual computing resource pool management functions mainly include dynamic resource configuration, quantitative evaluation of computing power, elastic expansion of computing power, computing power load balancing, and real-time monitoring of resources.

[0008] Dynamic resource configuration: Establish dynamic resource management, abstract different types and manufacturers of edge-side devices into different virtual resources according to their processing capabilities, and configure computing resources (GPU, NPU, TPU and VPU), network resources (Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485), and storage resources (physical storage, virtual storage, shared storage) on different virtual resources according to the physical connection method.

[0009] Computing power quantitative evaluation: Establish a computing power quantitative evaluation mechanism to conduct a comprehensive evaluation of the number of floating-point operations per second, the number of instructions executed per second, and the memory parameters of a single computing power resource, and provide computing power quantitative indicators to provide basic data for computing power scheduling and allocation.

[0010] Elastic expansion of computing power: Supports dynamic expansion of heterogeneous computing resources, including dynamic addition and dynamic unloading of computing resources.

[0011] Real-time resource monitoring: including statistics on the actual occupancy rate of computing resources (GPU, NPU, TPU and VPU), network resources (Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485) and storage resources (physical storage, virtual storage, shared storage) related to computing applications on virtual resources, to analyze the current remaining computing processing capacity and the number of computing tasks that can be undertaken in the system in the future.

[0012] Computing load balancing: supports load balancing of heterogeneous computing resources. When the computing resource load in the system is too high or too low, timely coordinate the distribution of related computing tasks on computing resources to balance the load of computing tasks on various computing resources.

[0013] The cross-domain distributed computing power soft bus part, based on the soft bus technology, realizes the cross-domain call of distributed computing power. By establishing intermediate routing forwarding nodes for computing power resources between different communication domains, the routing forwarding nodes support common power system physical layer protocols such as Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485, and real-time synchronous update of the computing power resource list between different communication protocol networks. Based on the soft bus technology, a real-time synchronization mechanism for cross-domain computing power resources is implemented to realize the subscription, release and remote call of various computing power resources across domains. Supports the default computing power soft bus routing configuration. When computing power resources cannot be obtained in a communication domain, the specified default soft bus is routed to the forwarding node, and the globally available idle computing power resource list can be queried across communication domains. Supports the soft bus forwarding node to regularly synchronize and update the computing power resource routing table of other communication networks across communication domains on the soft bus, so that the computing power resources between different communication networks are visible to each other. The computing resource allocation strategy is optimized through intelligent routing methods. By evaluating the relationship between computing tasks and computing resource requirements, computing resources are selected and computing routing allocation is adjusted in real time according to the current computing power status, real-time load, and network status to achieve optimal configuration of computing resource utilization. The computing power soft bus routing QoS is supported. When there are multiple intermediate routing forwarding nodes between two communication networks in the system, the appropriate routing forwarding node is selected according to the current load of the forwarding node and the communication bandwidth capacity of different protocols to achieve cross-domain communication mode optimization.

[0014] The functions of the cross-domain distributed computing power soft bus mainly include cross-domain computing power resource synchronization, computing power soft bus routing, cross-domain computing power soft bus routing QoS, and computing power resource intelligent routing.

[0015] Cross-domain computing power resource synchronization: Based on soft bus technology, a real-time synchronization mechanism for cross-domain computing power resources is implemented to realize the subscription, release and remote call of various cross-domain computing power resources.

[0016] Computing power soft bus routing: supports computing power soft bus routing configuration. When computing power resources cannot be obtained within a communication domain, they are routed to the forwarding node through the specified default soft bus. The list of globally available idle computing power resources can be queried across communication domains.

[0017] Intelligent routing of computing resources: Optimize computing resource allocation strategies through intelligent routing methods. By evaluating the relationship between computing tasks and computing resource requirements, select computing resources and adjust computing routing allocation in real time according to the current computing status, real-time load, and network status, and achieve optimal configuration of computing resource utilization. Support computing soft bus routing QoS. When there are multiple intermediate routing forwarding nodes or multiple communication channels between two communication networks in the system, select appropriate routing forwarding nodes and communication channels according to the current load of the forwarding nodes and the bandwidth capacity of different communication channel protocols to achieve cross-domain communication mode optimization.

[0018] The dynamic computing power scheduling part includes a dynamic computing power scheduling system based on the computing power resource pool to realize the dynamic allocation and management of computing power resources. Business applications do not need to consider the nodes where the physical computing resources of the algorithm are finally executed. The computing power requirements of each computing task are evaluated in real time, and the virtual computing resource pool is requested to allocate the best computing power resource combination for the algorithm application. When allocating computing power resources for algorithm computing tasks, the status of the computing power resource pool is first obtained, and the status of computing resources (computing power, network load, availability, etc.), task priority, data dependency between tasks, cross-domain communication delay and other factors are comprehensively considered to dynamically obtain heterogeneous computing resources. For high-reliability algorithm computing tasks, the computing power scheduling system monitors the health status of the virtual computing resource pool in real time. Once an abnormality is found in the computing resources being executed, the computing tasks on this computing resource are promptly transferred to other normal computing resources for execution to ensure the integrity of the execution results. Provide a unified algorithm call interface service, provide an algorithm call interface for each application, distribute algorithm computing tasks to the computing resource pool, shield the physical heterogeneity characteristics of the computing power nodes that actually carry the computing tasks, and for the business, the implementation of the application algorithm is independent of the location of the actual computing power running node.

[0019] The functions of the dynamic computing power scheduling system mainly include computing power demand assessment, task computing power scheduling, computing power task transfer, etc.

[0020] Computing power demand assessment: Evaluate the computing power requirements of each computing task in real time, and request the virtual computing resource pool to allocate the best computing power resource combination for the algorithm application.

[0021] Task computing power scheduling: When allocating computing power resources for algorithm computing tasks, first obtain the status of the computing power resource pool, comprehensively consider multiple factors such as the status of computing resources (computing power, network load, availability, etc.), task priority, data dependencies between tasks, cross-domain communication delays, and dynamically obtain heterogeneous computing resources.

[0022] Computing task transfer: For high-reliability algorithm computing tasks, the computing scheduling system monitors the health status of the virtual computing resource pool in real time. Once an abnormality is found in the computing resources being executed, the computing task on the current computing resource will be transferred to other normal computing resources in a timely manner to ensure the integrity of the execution results.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: the edge-side heterogeneous computing power resource scheduling provided by the present invention can be widely used in intelligent scenario applications of various edge-side devices in new power systems, such as integrated terminals, concentrators, station house gateways, and downstream devices such as inverters, charging piles, energy storage, environmental monitoring, video monitoring, energy meters and other equipment systems. Through cloud-edge and edge-edge collaboration, the algorithm tasks that need to be executed are dynamically configured for edge-side computing power resources, increasing the intelligent application collaboration capabilities of the edge-side system. Through unified management of computing power resources and real-time cross-domain computing power soft bus, this system supports the system's support for computing power resource scheduling, dynamic balanced migration and other functions under heterogeneous asymmetric computing power architecture, solves the problems of difficult resource sharing and low efficiency of computing power collaborative scheduling after the introduction of a large number of heterogeneous devices and computing power resources in new power systems, reduces the management cost of heterogeneous computing power resources on the edge, and improves the overall utilization efficiency of computing power resources in the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the system deployment structure for scheduling heterogeneous computing power resources on the edge side and the cloud side provided by the present invention; Figure 2 It is a schematic diagram of the management structure of the heterogeneous computing power virtual computing resource pool provided by the present invention; Figure 3 A basic flow chart of a cross-domain distributed computing power soft bus application according to an embodiment of the present invention; Figure 4 This is a basic flow chart of the dynamic computing power scheduling system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. Example

[0026] The system application structure in Example 1 of the present invention can be divided into the cloud side and the edge side, where the cloud side application mainly implements cloud-edge instructions to call the computing resources on the edge side, and the edge side application calls all available computing resources in the communication domain and all available computing resources in the communication domain of other edge devices with communication connections through instructions.

[0027] In the heterogeneous computing power virtual computing resource pool management structure in Example 2 of the present invention, the virtual computing resource pool includes all computing resources of the cloud edge, including the central processing unit CPU, the graphics processing unit GPU, the video processing unit VPU, the neural network processor NPU, etc. The heterogeneous computing power virtual computing resource pool uniformly manages and virtualizes various device systems on the edge side according to computing power resources, network resources and storage resources to form a unified virtual computing resource pool, which is uniformly scheduled by the computing power scheduling module on the edge side. Different types and manufacturers of edge side devices are abstracted into different virtual computing resources according to their processing capabilities, and computing power resources (GPU, NPU, TPU and VPU), network resources (Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485), and storage resources (physical storage, virtual storage, shared storage) are configured on different virtual computing resources according to the physical connection method to achieve elastic management, thereby supporting dynamic configuration, elastic expansion, load balancing, fault detection and other mechanisms of computing power resources to ensure high availability and operational reliability of computing power resources during the execution of computing tasks. The method includes: The edge-side computing power scheduling 201 provided in the embodiment of the present invention uniformly manages the edge-side computing power resource pool and completes the unified scheduling of edge-side computing power resources.

[0028] The edge computing power management node 202 provided in the embodiment of the present invention includes various intelligent management terminals deployed on the edge, which are responsible for managing the computing power resources of each end side connected within the communication domain.

[0029] The end-side computing power resources 203 provided in this embodiment of the present invention include end-side computing power devices such as GPU, NPU, TPU and VPU.

[0030] Figure 3 This is a basic flow chart of the cross-domain distributed computing power soft bus application of an embodiment of the present invention. The present invention uses this to illustrate a method for self-networking and communication of computing power resources of heterogeneous devices on the edge side through a soft bus. In the figure, the communication domain A, the fusion terminal and the new energy monitoring equipment, the communication domain B, the communication domain A and the communication domain B, the fusion terminal and the artificial intelligence terminal, all use one or more heterogeneous networking methods such as Ethernet, Bluetooth, WIFI, HPLC, CAN, LORA, NBIOT, 4G, high-speed RS485, etc. to communicate, and the cross-domain call of distributed computing power resources is realized through the computing power soft bus. The method includes: In the cross-domain computing power resource synchronization 301 provided by the embodiment of the present invention, both the fusion terminal and the artificial intelligence terminal belong to the intermediate routing forwarding node, and the real-time synchronization of computing power resources across communication domain A and communication domain B is realized through the soft bus technology, so that the fusion terminal can quickly identify the computing power equipment in communication domain B, such as the VPU computing power resources on video surveillance 1 and video surveillance 2, and realize the synchronization and self-discovery of computing power resources between different communication networks. The artificial intelligence terminal periodically synchronizes the current computing power status, real-time load, network status, etc. of all computing power resources (VPU) in the communication domain, so that the fusion terminal can grasp the actual status of all current computing power resources.

[0031] The computing power soft bus routing table 302 provided in the embodiment of the present invention supports the computing power soft bus forwarding node to periodically and synchronously update the computing power resource routing table of other communication networks across communication domains, so that the computing power resources between different communication networks are mutually visible. When the fusion terminal cannot obtain computing power resources in the communication domain A, it queries the globally available idle computing power resource list through the computing power resource routing table, finds the computing power routing forwarding node such as the artificial intelligence terminal, and can select the appropriate computing power routing forwarding node when performing algorithm calculation, thereby realizing the remote call of cross-domain computing power resources.

[0032] The computing power resource intelligent routing selection 303 provided in the embodiment of the present invention, the fusion terminal optimizes the computing power resource allocation strategy through the intelligent routing method, and selects computing power resources and adjusts computing power routing allocation in real time according to the current computing power status, real-time load, and network status of various computing power devices on the communication domain B by evaluating the relationship between computing tasks and computing power resource requirements, so as to achieve the optimal configuration of computing power resource utilization. Support computing power soft bus routing QoS. When there are multiple intermediate routing forwarding channels or multi-level forwarding nodes between communication domain A and communication domain B in the system, according to the current load of the forwarding node and the bandwidth capacity of different communication channel protocols, select the appropriate routing forwarding node and communication mode to achieve cross-domain communication mode optimization. For example, when the fusion terminal and the artificial intelligence terminal have both Ethernet and RS485 communication channels, the Ethernet communication mode with higher performance is generally selected.

[0033] Figure 4 This is a basic flow chart of the computing power dynamic scheduling system of an embodiment of the present invention, and the present invention uses this to illustrate a process and method for performing algorithm calculations by dynamically scheduling edge-side computing power resources.

[0034] The computing power requirement assessment 401 provided in the embodiment of the present invention is that the fusion terminal accepts computing tasks, evaluates the computing power requirement required for each computing task in real time, and determines the amount of virtual computing power resources that the virtual computing resource pool needs to allocate for the algorithm application.

[0035] The computing power resource allocation 402 provided in the embodiment of the present invention, before the fusion terminal allocates computing power resources to the algorithm computing task, firstly, the computing power resources synchronously obtain the status of the computing resources (computing power, network load, availability, etc.), comprehensively considers multiple factors such as the priority of the current computing task, the data dependency relationship between tasks, cross-domain communication delay, etc., and analyzes and determines the optimal actual computing power resource combination.

[0036] The computing power task scheduling 403 provided in the embodiment of the present invention is that the computing power scheduling module analyzes the computing power resources and determines the actual computing power resource combination. If the computing power resources are within the local communication network, they are directly called; if the computing power resources need to cross domains, the computing power resources are called across domains through the computing power routing node.

[0037] The computing power task transfer 404 provided in the embodiment of the present invention, the computing power scheduling module will monitor the health status of the computing resources in real time for high-reliability algorithm computing tasks. Once an abnormality is found in the computing resources being executed, the computing power will be re-scheduled to transfer the computing tasks on the current computing resources to other normal computing resources in time for execution, so as to ensure the integrity of the execution results.

[0038] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A heterogeneous computing resource scheduling system on the edge side, characterized in that: The system includes a computing power virtualization resource pool module, a cross-domain distributed computing power soft bus, and a computing power resource scheduling management module, wherein: The computing power virtualization resource pool module, based on the heterogeneous computing power virtual computing resource pool technology, uniformly manages and virtualizes various device systems on the edge side according to computing power resources, network resources and storage resources to form a unified virtual computing resource pool; The cross-domain distributed computing power soft bus establishes intermediate routing and forwarding nodes for computing power resources between different communication domains. The routing and forwarding nodes support the physical layer protocol of the power system and synchronously update the computing power resource lists between networks with different communication protocols in real time. The computing power resource scheduling management module is a dynamic computing power scheduling system established based on the computing power resource pool. It realizes the dynamic allocation management of computing power resources, evaluates the computing power requirements of each computing task in real time, and requests the virtual computing resource pool to allocate the best computing power resource combination for the algorithm application.

2. According to claim 1, a heterogeneous computing resource scheduling system on the edge side is characterized in that: The computing power virtualization resource pool module abstracts different types and manufacturers of edge-side equipment into different virtual computing resources according to their processing capabilities, configures computing power resources, network resources, and storage resources on different virtual computing resources based on physical connection methods, and quantifies their computing power indicators to carry out different computing power tasks.

3. The edge-side heterogeneous computing resource scheduling system according to claim 1, characterized in that: The cross-domain distributed computing power soft bus, when it is impossible to obtain computing power resources within a communication domain, routes to the forwarding node through the specified default soft bus, can query the list of globally available idle computing power resources across communication domains, optimize the computing power resource allocation strategy through an intelligent routing method, and select computing power resources and adjust computing power routing allocation in real time according to the current computing power status, real-time load, and network status by evaluating the relationship between computing tasks and computing power resource requirements, thereby achieving optimal configuration of computing power resource utilization.

4. The edge-side heterogeneous computing resource scheduling system according to claim 1, characterized in that: When allocating computing resources to algorithm computing tasks, the computing resource scheduling management module first obtains the status of the computing resource pool, comprehensively considers the status of computing resources, task priorities, data dependencies between tasks, and cross-domain communication delay factors, and dynamically obtains heterogeneous computing resources.

5. The edge-side heterogeneous computing resource scheduling system according to claim 4, characterized in that: For high-reliability algorithm computing tasks, the computing power resource scheduling management module monitors the health status of the virtual computing resource pool in real time. Once an abnormality is found in the computing resources being executed, the computing tasks on the current computing resources are promptly transferred to other normal computing resources for execution to ensure the integrity of the execution results. A unified algorithm calling interface service is provided, an algorithm calling interface is provided for each application, and algorithm computing tasks are distributed to the computing resource pool, shielding the internal physical heterogeneous characteristics of the computing power nodes that actually carry the computing tasks.

6. The edge-side heterogeneous computing resource scheduling system according to claim 3, characterized in that: When there are multiple intermediate routing forwarding nodes between two communication networks in the system, a suitable routing forwarding node is selected according to the current load of the forwarding node and the communication bandwidth capacity of different protocols.

7. The edge-side heterogeneous computing resource scheduling system according to claim 2, characterized in that: The computing power virtualization resource pool module includes computing power load balancing, including load balancing that supports heterogeneous computing power resources. When the computing power resource load in the system is too high or too low, it promptly coordinates the distribution of related computing tasks on the computing resources to balance the load of computing power tasks on each computing resource.

8. The edge-side heterogeneous computing resource scheduling system according to claim 2, characterized in that: The computing power virtualization resource pool module includes computing power quantitative evaluation, including establishing a computing power quantitative evaluation mechanism, conducting a comprehensive evaluation on the number of floating-point operations per second, the number of instructions executed per second and the video memory parameters of a single computing power resource, and providing computing power quantitative indicators to provide basic data for computing power scheduling and allocation.

9. The edge-side heterogeneous computing resource scheduling system according to claim 2, characterized in that: The computing power virtualization resource pool module includes real-time resource monitoring, including statistics on the actual occupancy rate of computing power resources, network resources, and storage resources related to computing power applications on virtual resources, which is used to analyze the current remaining computing processing capacity and the number of computing power tasks that can be undertaken in the system in the future.

10. A method for scheduling heterogeneous computing resources on an edge side, characterized in that: The method comprises the following steps, Step S1, computing power demand assessment, the fusion terminal accepts the computing task, evaluates the computing power demand required for each computing task in real time, and determines the amount of virtual computing power resources that the virtual computing resource pool needs to allocate for the algorithm application; Step S2, computing power resource allocation. Before the fusion terminal allocates computing power resources for the algorithm computing task, the computing power resources first synchronously obtain the computing resource status, comprehensively consider the priority of the current computing task, the data dependency relationship between tasks, cross-domain communication delay and other factors, and analyze and determine the optimal actual computing power resource combination; Step S3, computing task scheduling, the computing scheduling module analyzes the computing resources and determines the actual computing resource combination. If the computing resources are in the local communication network, they are directly called; if the computing resources need to cross domains, the computing resources are called across domains through the computing routing node; Step S4, computing power task transfer. For high-reliability algorithm computing tasks, the computing power scheduling module will monitor the health status of computing resources in real time. Once an abnormality is found in the computing resources being executed, the computing power will be re-scheduled to transfer the computing tasks on the current computing resources to other normal computing resources in time to ensure the integrity of the execution results.

Images