Computing power sensing method and system for router, gateway, IPC and ONU

Abstract

The invention provides a computing power sensing method and system for a router, a gateway, an IPC and an ONU, and belongs to the technical field of computing power sensing. The method comprises the steps that S1, networking is carried out after a computing power node is started; s2, each computing power node detects network resources, computes computing resources of the node and further creates a service state information table; s3, each computing power node performs computing power evaluation on the computing power request based on the service state information table of the node to generate a computing power evaluation result; step S4, based on the network bandwidth, the network rate, the data flow viscosity and the dynamic anycast availability, calculating the scheduling priority of the other computing power nodes, and forwarding the computing power request based on the scheduling priority and the computing power evaluation result; and S5, the computing power node receives the forwarded computing power request through the computing power entry, calls the computing power service based on the computing power request, and outputs an execution result through the computing power exit. The method has the advantages that the load balancing capacity and expansibility of the computing power node are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power perception, and particularly to a method and system for computing power perception for routers, gateways, IP cameras, and ONUs. Background Art

[0002] Computing-Networking Perception involves the collaborative work of computing resources and network resources, and requires real-time perception and intelligent scheduling of the resource status of these computing resources and network resources; Computing-Networking Perception is reflected in Computing-Aware Networking (abbreviated as CAN). Especially in the context of the wide application of big data, 5G, and edge computing, the function of Computing-Networking Perception is more worthy of exploration.

[0003] Currently, the research and development of computing-aware networking are in the early stage, mainly relying on existing network management and automation protocols, such as BGP (Border Gateway Protocol) and cloud-native technologies. Container technology (such as Docker) is used to package the applications and their dependencies of each node (such as routers, gateways, IP cameras (network cameras), ONUs, etc.) located in the edge network. Each node is an independent computing power perception node. By using declarative APIs (such as the declarative API of Kubernetes), the status of applications and resources in each node is managed. The data running on the node is carried in the native protocol for transmission. The dynamic deployment, expansion, and management of nodes are usually achieved through container orchestration tools (such as Kubernetes).

[0004] However, the traditional method has the following disadvantages: 1. Using BGP to implement computing-aware networking lacks customization of computing power. It only simply performs computing power boundary perception through network exploration. The perception information transmission and abstraction are limited, and the purpose of free expansion cannot be achieved; 2. Cloud-native technologies can perform custom encapsulation and information transmission through remote tools and declarative APIs. However, the native data interfaces and APIs follow a set of frameworks, with insufficient freedom and being too bloated. It is impossible to perform specific data serialization (that is, to uniformly package all data into data packets and unable to quickly obtain the required data from the data packets), and it is impossible to converge, synchronize, and heterogeneousize node information (that is, unable to uniformly manage nodes with different architectures and different algorithms). There is still room for improvement in custom protocol extension.

[0005] Due to the shortcomings of traditional methods, the computing power of each node in the computing power perception network cannot be well load-balanced, and the scalability is poor, which in turn affects the computing efficiency and cost, especially the computing of edge nodes located in the edge network. Therefore, how to provide a computing power perception method and system for routers, gateways, IPCs, and ONUs to improve the load-balancing ability and scalability of computing power nodes has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a computing power perception method and system for routers, gateways, IPCs, and ONUs to improve the load-balancing ability and scalability of computing power nodes.

[0007] In a first aspect, the present invention provides a computing power perception method for routers, gateways, IPCs, and ONUs, including the following steps:

[0008] Step S1: After a computing power node of which the device type is a physical device or a virtual device is started, it forms a computing power network based on the SDN protocol; the physical device is a router, a gateway, an IPC, or an ONU;

[0009] Step S2: Each computing power node in the computing power network starts a network detection thread, detects network resources through the network detection thread, and stores the network resources;

[0010] Step S3: Each computing power node calculates the computing resources of its own node in real time and stores the computing resources;

[0011] Step S4: Each computing power node creates a service status information table based on the network resources and computing resources for computing power perception;

[0012] Step S5: When each computing power node receives a computing power request, it performs a computing power evaluation on the computing power request based on the service status information table of its own node to generate a computing power evaluation result;

[0013] Step S6: Each computing power node calculates the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow viscosity, and dynamic anycast availability, and forwards the computing power request based on the scheduling priorities and the computing power evaluation result;

[0014] Step S7: The computing power node receives the forwarded computing power request through the computing power entrance, calls the corresponding computing power service based on the computing power request and executes it to generate an execution result, and outputs the execution result through the computing power exit.

[0015] Further, step S1 is specifically:

[0016] After a computing power node with a device type of physical device or virtual device is started, through a computing power adapter, it obtains the broadcast data packets broadcast by other computing power nodes based on the SDN protocol, establishes a two-way connection with the corresponding computing power nodes based on the broadcast data packets, and exchanges the node information of the computing power nodes, thereby forming a computing power network; the physical device is a router, gateway, IPC or ONU; the virtual device is deployed on a server;

[0017] Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

[0018] Further, the specific steps of step S2 are as follows:

[0019] When each computing power node in the computing power network is started, a network detection thread is started synchronously respectively. Through the network detection thread, network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput, and jitter are detected in real time, and the network resources are stored;

[0020] Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay;

[0021] The specific steps of step S3 are as follows:

[0022] Each computing power node calculates in real time the computing resources of this node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate, and load information, and stores the computing resources;

[0023] The load information includes at least the number of sessions being served, the number of queries per second, and the calculation delay.

[0024] Further, the specific steps of step S4 are as follows:

[0025] Each computing power node creates a service status information table including a routing sub-table based on the network resources and computing resources for computing power perception, and shares the service status information table to the computing power network through the SDN protocol;

[0026] The specific steps of step S5 are as follows:

[0027] When each computing power node receives a computing power request carrying at least a computing task, a computing time limit, and a computing priority, it performs a computing power evaluation on the computing power request based on the service status information table of this node, and judges whether the network resources and computing resources of this node can meet the computing task carried by the computing power request, thereby generating a computing power evaluation result of sufficient computing power or insufficient computing power.

[0028] Further, the specific steps of step S6 are as follows:

[0029] Each of the computing power nodes calculates the weights of the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability of the remaining computing power nodes respectively, performs weighted summation on the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability based on the weights, and then calculates the scheduling priorities of the remaining computing power nodes;

[0030] When the computing power evaluation result indicates sufficient computing power, the corresponding computing power service is invoked through the computing power entry of this node to execute the computing power request, and an execution result is generated, which is output through the computing power exit of this node; when the computing power evaluation result indicates insufficient computing power, the computing power request is forwarded based on the scheduling priority;

[0031] In step S7, both the computing power entry and the computing power exit are API interfaces, and the call status of the computing power entry and the computing power exit is fed back through the status callback interface;

[0032] The computing power entry is used for real-time addressing and computing power scheduling of the computing power service corresponding to the computing power request; the computing power exit is used for processing query, aggregation, output, and full-network forwarding of the execution result;

[0033] Each of the computing power services has a unique service ID, representing an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

[0034] In a second aspect, the present invention provides a computing power perception system for routers, gateways, IP cameras, and ONUs, including the following modules:

[0035] The computing power node networking module is used for, after the computing power nodes of physical devices or virtual devices are started, networking based on the SDN protocol to form a computing power network; the physical devices are routers, gateways, IP cameras, or ONUs;

[0036] The network resource detection module is used for each computing power node in the computing power network to start a network detection thread respectively, detect network resources through the network detection thread, and store the network resources;

[0037] The computing resource calculation module is used for each computing power node to calculate the computing resources of this node in real time and store the computing resources;

[0038] The service status information table creation module is used for each computing power node to create a service status information table based on the network resources and computing resources for computing power perception;

[0039] The computing power evaluation module is used for each computing power node to perform computing power evaluation on the computing power request based on the service status information table of this node when receiving a computing power request, and generate a computing power evaluation result;

[0040] A load balancing module, configured to enable each of the computing power nodes to calculate the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow viscosity, and dynamic anycast availability, and forward the computing power requests based on the scheduling priorities and the computing power evaluation results;

[0041] A computing power request execution module, configured to enable the computing power node to receive the forwarded computing power request through the computing power entrance, call the corresponding computing power service based on the computing power request and execute it, generate an execution result, and output the execution result through the computing power exit.

[0042] Further, the computing power node networking module is specifically configured to:

[0043] After a computing power node with a device type of a physical device or a virtual device is started, through a computing power adapter, based on the SDN protocol, obtain broadcast data packets broadcast by the remaining computing power nodes, establish a two-way connection with the corresponding computing power nodes based on the broadcast data packets, and exchange the node information of the computing power nodes, thereby forming a computing power network; the physical device is a router, a gateway, an IPC, or an ONU; the virtual device is deployed on a server;

[0044] Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

[0045] Further, the network resource detection module is specifically configured to:

[0046] When each computing power node in the computing power network is started, a network detection thread is synchronously started respectively, and the network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput, and jitter are detected in real time through the network detection thread, and the network resources are stored;

[0047] Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay;

[0048] The computing resource calculation module is specifically configured to:

[0049] Each computing power node calculates in real time the computing resources of this node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate, and load information, and stores the computing resources;

[0050] The load information includes at least the number of sessions being served, the number of queries per second, and the computing delay.

[0051] Further, the service status information table creation module is specifically configured to:

[0052] Each of the computing power nodes creates a service status information table including a routing sub-table based on the network resources and computing resources for computing power perception, and shares the service status information table to the computing power network through the SDN protocol;

[0053] The computing power evaluation module is specifically used for:

[0054] When each of the computing power nodes receives a computing power request carrying at least a computing task, a computing time limit, and a computing priority, it performs computing power evaluation on the computing power request based on the service status information table of this node, and determines whether the network resources and computing resources of this node can meet the computing task carried by the computing power request, and then generates a computing power evaluation result of sufficient computing power or insufficient computing power.

[0055] Further, the load balancing module is specifically used for:

[0056] Each of the computing power nodes calculates the weights of the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability of the remaining computing power nodes respectively, performs weighted summation on the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability based on the weights, and then calculates the scheduling priorities of the remaining computing power nodes;

[0057] When the computing power evaluation result is sufficient computing power, call the corresponding computing power service through the computing power entry of this node to execute the computing power request, generate an execution result, and output the execution result through the computing power exit of this node; when the computing power evaluation result is insufficient computing power, forward the computing power request based on the scheduling priority;

[0058] In the computing power request execution module, both the computing power entry and the computing power exit are API interfaces, and the call status of the computing power entry and the computing power exit is fed back through the status callback interface;

[0059] The computing power entry is used for real-time addressing and computing power scheduling of the computing power service corresponding to the computing power request; the computing power exit is used for processing query, aggregation, output, and full-network forwarding of the execution result;

[0060] Each of the computing power services is provided with a unique service ID, representing an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

[0061] The advantages of the present invention are:

[0062] 1. After the computing power nodes are started, a computing power network is formed through networking based on the SDN protocol. Each computing power node in the computing power network starts a network detection thread respectively, detects network resources through the network detection thread and stores them. Each of the computing power nodes calculates and stores its own computing resources in real time, creates a service status information table based on the network resources and computing resources for computing power perception. Then, when each computing power node receives a computing power request, it performs computing power evaluation on the computing power request based on the service status information table of its own node to generate a computing power evaluation result. Then, each computing power node calculates the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow viscosity, and dynamic anycast availability, and forwards the computing power request based on the scheduling priorities and the computing power evaluation result. Finally, the computing power node receives the forwarded computing power request through the computing power entrance, invokes the corresponding computing power service based on the computing power request and executes it, and outputs the execution result through the computing power exit. That is, based on the SDN protocol, computing power nodes with different architectures and different algorithms are connected, node information is aggregated, synchronized, and heterogeneousized, network resources and computing resources are dynamically and real-time perceived to create a service status information table, and then the computing tasks carried by the computing power requests are overallocated and scheduled, forming a computing power network in which computing power resources (network resources and computing resources) can be perceived, allocated, and scheduled. When the computing power of a certain computing power node is insufficient, the computing power of other computing power nodes is borrowed for load balancing. And computing power scheduling is performed through the service status information table, that is, for specific data serialization, the service status information table can be flexibly defined to meet different expansion requirements, ultimately greatly improving the load balancing ability and scalability of the computing power nodes, thereby greatly improving the computing efficiency of edge computing and greatly reducing the computing cost of edge computing.

[0063] 2. Through a lightweight SDN protocol for network exploration, data encapsulation, data reporting, and data synchronization, the status of network resources and computing resources is monitored and perceived in real time, and reported to the edge layer proximal in the computing power network to push the computing task to the network edge of the computing power network, making the computing task closer to the data source or user, effectively reducing transmission latency, and ensuring service quality (QoS) and user experience (QoE), especially in real-time and bandwidth-sensitive applications, achieving the goal of computing power load balancing and realizing a certain scale of breakthrough in computing power efficiency.

[0064] 3. A computing power network is formed through the SDN (software-defined networking) protocol. Since the SDN protocol can program the control plane, the separation of the data plane and the control plane will help abstract the underlying network infrastructure resources and provide APIs in the form of programmable network objects. By programming network objects, there is no need to focus on the specific logic of underlying network devices such as routers and switches, which has better flexibility and controllability. Moreover, the SDN protocol design can focus on the content and forwarding logic of data without considering the underlying decision-making strategies. All decision-making strategies are determined and processed by the SDN controller on the control plane, which makes the configuration and management of the computing power network very agile and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.

[0066] Figure 1 is a flowchart of a computing power perception method for routers, gateways, IP cameras, and ONUs according to the present invention.

[0067] Figure 2 is a schematic structural diagram of a computing power perception system for routers, gateways, IP cameras, and ONUs according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The overall idea of the technical solution in the embodiments of the present application is as follows: Based on the SDN protocol, distributed computing power nodes with different architectures and different algorithms are connected, and network resources and computing resources are dynamically and real-time sensed to create a service status information table. Then, the computing tasks carried by the computing power requests are overallocated and scheduled to form a computing power network in which computing power resources can be sensed, allocated, and scheduled. Moreover, through the service status information table for computing power scheduling, the service status information table can be flexibly defined to meet different expansion requirements, thereby improving the load balancing ability and scalability of the computing power nodes.

[0069] Please refer to Figures 1 to 2 As shown, a preferred embodiment of a computing power perception method for routers, gateways, IP cameras, and ONUs according to the present invention includes the following steps:

[0070] Step S1. After a computing power node of which the device type is a physical device or a virtual device is started, networking is performed based on the SDN protocol to form a computing power network; the physical device is a router, a gateway, an IP camera, or an ONU.

[0071] Step S2. Each computing power node in the computing power network starts a network detection thread, detects network resources through the network detection thread, and stores the network resources.

[0072] Step S3: Each computing power node calculates the computing resources of its own node in real time and stores the computing resources.

[0073] Step S4: Each computing power node creates a service status information table based on the network resources and computing resources for computing power perception.

[0074] Step S5: When each computing power node receives a computing power request, it performs a computing power evaluation on the computing power request based on the service status information table of its own node to generate a computing power evaluation result.

[0075] Step S6: Each computing power node calculates the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow viscosity, and dynamic anycast availability, and forwards the computing power request based on the scheduling priorities and the computing power evaluation result.

[0076] Step S7: The computing power node receives the forwarded computing power request through the computing power entrance, invokes the corresponding computing power service based on the computing power request and executes it to generate an execution result, and outputs the execution result through the computing power exit.

[0077] Specifically, Step S1 is as follows:

[0078] After a computing power node with a device type of physical device or virtual device is started, through a computing power adapter, based on the SDN protocol, it obtains the broadcast data packets broadcast by the remaining computing power nodes, establishes a two-way connection with the corresponding computing power nodes based on the broadcast data packets, and exchanges the node information of the computing power nodes, thereby forming a computing power network; the physical device is a router, gateway, IPC, or ONU; the virtual device is deployed on a server; the broadcast data packet carries at least a preamble, access address, protocol data unit, CRC check code, device name, service UUID, and flag bit.

[0079] Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

[0080] The computing power node is a basic functional entity in the computing power network, usually integrated in an edge computing node, can provide the load situation of the computing resources of its own node in real time, and can also provide computing power services to other computing power nodes; when there is a computing power request in the entire computing power network, it can be scheduled and allocated through the computing power resource pool (edge computing power resource pool / cloud computing power resource pool).

[0081] The Computing Power Adapter is an important functional entity in the computing power network, mainly used to help terminal devices (computing power nodes) access the computing power network and achieve efficient scheduling and management of computing power resources. The computing power adapter helps terminal devices access the computing power network by maintaining bound IP (BIP) information, identifying initial request packets, etc. The computing power adapter can be implemented as part of the computing power network (internal mode) or deployed on a separate device (external mode), such as a user plane function (UPF) in a mobile network or a broadband remote access server (BRAS) in a fixed network. In the external mode, the computing power adapter can be placed closer to the terminal device, while the computing power nodes can be placed at an aggregation point and connected to multiple computing power adapters; compared with the internal mode, the external computing power adapter can retain less client binding information, and the functional processing is relatively independent with weak coupling.

[0082] The specific steps of step S2 are as follows:

[0083] When each computing power node in the computing power network starts up, a network detection thread is synchronously started respectively. Through the network detection thread, network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput, and jitter are detected in real time, and the network resources are stored.

[0084] Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay. Specifically, in implementation, it can be calculated based on average calculation or norm calculation to improve the reliability of jitter calculation; the two-dimensional array of the transmission delay records the node end point (the computing power node at the transmission end) and the round-trip response time, and the transmission delay within a certain period of time is stored by accumulating the two-dimensional array per second; the network bandwidth can be detected by downloading model data, that is, by downloading model data with a large amount of data for speed measurement.

[0085] The specific steps of step S3 are as follows:

[0086] Each computing power node calculates in real time the computing resources of this node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate, and load information, and stores the computing resources.

[0087] The load information includes at least the number of sessions being served, the number of queries per second, and the computing delay.

[0088] The specific steps of step S4 are as follows:

[0089] Each computing power node creates a service status information table containing a routing sub-table based on the network resources and computing resources for computing power perception, and shares the service status information table to the computing power network through the SDN protocol.

[0090] The service status information table is updated and synchronized regularly. To avoid the impact of network fluctuations, the update and synchronization of the service status information table of the computing power nodes are only performed when the metric changes exceed the threshold or the update timer expires.

[0091] The specific steps of step S5 are as follows:

[0092] When each computing power node receives a computing power request carrying at least a computing task, a computing time limit, and a computing priority, it performs a computing power evaluation on the computing power request based on the service status information table of this node, and determines whether the network resources and computing resources of this node can meet the computing task carried by the computing power request, and then generates a computing power evaluation result of sufficient computing power or insufficient computing power.

[0093] The specific steps of step S6 are as follows:

[0094] Each computing power node calculates the weights of the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability of the remaining computing power nodes respectively, performs a weighted sum on the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability based on the weights, and then calculates the scheduling priority of the remaining computing power nodes;

[0095] The specific process of calculating the weight of the network bandwidth is as follows: for network types such as 5G networks and Wi-Fi networks, the computing power node jumps to the transmission delay. Assume that the transmission delay from point A to point B is 5ms, and the data storage is AB5. After merging the vertices, the transmission delay between any two connectable computing power nodes is obtained through the shortest path algorithm. As long as it is judged that the transmission delay of the same vertex in the database is relatively large, it can be replaced; in the initial state, the states of all computing power nodes are the same, and only the transmission delay is used to increase the weight. In short, there are multiple paths connecting point A and point B, and the corresponding weights are set based on the transmission delays of each path.

[0096] The specific process of calculating the weight of the network rate is as follows: The same template data is sent to two adjacent computing power nodes. Set the computing power node with slow efficiency as the base value 1, and the other computing power node is given a relative value (less than 1). Before requesting the template data, the computing power required for the request response is selected, and the relevant template data is pushed, and so on. After pairwise comparison, the relative value of the computing power delay is obtained, and the weight of the network rate is set based on this relative value.

[0097] In specific implementation, considering the routing selection mechanism of the optimal computing power node, the routing sub-table of the computing power node contains status information data of computing resources and network resources, including evaluation parameters of computing resource performance (CPU, GPU, NPU, memory, etc.), remaining computing capacity (if the computing power has multi-core and multi-thread processing capabilities, this item is enabled), computing latency (network rate), etc. By performing weighted comprehensive calculation on network resources and computing resources, the optimal computing power node is selected for routing forwarding and task execution. When the routing node receives the data packet of the computing task, it parses the task TaskID, computing task type, data flow stickiness requirement attribute, computing power requirement range, longest computing duration, end notification type, etc. of the data packet. Based on the pre-obtained correspondence between the computing task type and the computing performance of other computing power nodes, at least one other computing power node corresponding to the computing task type and its computing performance are determined. Based on the computing performance of other nodes and the network link status between this node and other computing power nodes, the target computing power node is determined after comprehensive consideration; the address of the target computing power node is the routing destination address of the data packet, and then the data packet is forwarded based on the target address.

[0098] The data flow stickiness and dynamic anycast availability are used to accelerate the efficiency of selection and scheduling. The dynamic anycast availability can be evaluated based on latency, bandwidth, packet rate, and packet loss rate.

[0099] Data flow stickiness refers to the business data relationship that needs to be called long-term and frequently. For the business with high data flow stickiness, high frequency, and small granularity and the computing power, the weight can be strengthened to avoid problems such as disconnection, packet loss, and traffic chaos. Therefore, the service can submit data flow stickiness requirements to the computing power node, including flow stickiness type, timeout time, frequency, etc., generate and write them into the data flow stickiness requirement data table and store and copy the replicas at the computing power edge perception layer.

[0100] Dynamic anycast realizes routing computing power requests to equivalent service instances (computing power services) by dynamically calculating network resources and computing resources. Traditional anycast is usually used for single-request, single-response communication. When the network state changes, the relative positions of computing power nodes change. In a computing power-aware network, multi-request, multi-response communication is required between terminal devices and computing power nodes. Therefore, in the data plane of a computing power-aware network, it is necessary to maintain data flow stickiness to ensure that computing power requests are processed by the same computing power node, and this computing power node is determined when the first anycast request is received at the entrance of the sensing layer. After listening to the connection, the service access of different terminal devices with the same service ID at the entrance of the sensing layer can be scheduled to different computing power exits. This feature is called dynamic anycast. Dynamic anycast receives requests asynchronously in multiple threads. According to the requirements of edge computing power, the number of thread pools is configured to avoid a large number of requests affecting the use of memory space. The anycast service table contains flows, service IDs, computing power entrances, computing power exits, etc., and is stored on the computing power service or computing power adapter in the sensing layer, usually less than 100. That is, dynamic anycast dynamically adjusts the optimal path within the anycast group by real-time monitoring of network status (such as latency, load, etc.). For example, the anycast protocol based on mobile agents reduces the resource consumption of global information synchronization through local information collection and dynamic update of the routing table1; in a wireless sensor network, the area-disjoint strategy combines the neighbor node influence factor to optimize the anycast path selection and reduce the transmission collision caused by hidden terminals.

[0101] When the computing power evaluation result is that the computing power is sufficient, the corresponding computing power service is called through the computing power entrance of this node to execute the computing power request, generate an execution result, and output the execution result through the computing power exit of this node; when the computing power evaluation result is that the computing power is insufficient, the computing power request is forwarded based on the scheduling priority.

[0102] In specific implementation, the computing power exit selected by the computing power entrance is not necessarily the one with the lowest load; the computing power request can also be sent to the computing power exit with relatively lower computing load to avoid network fluctuations.

[0103] In step S7, both the computing power entrance and the computing power exit are API interfaces, and the call status of the computing power entrance and the computing power exit is fed back through the status callback interface.

[0104] The computing power entrance is used for real-time addressing and computing power scheduling of the computing power service corresponding to the computing power request; the computing power exit is used for querying, aggregating, outputting, and forwarding the execution result across the network.

[0105] Each of the computing power services has a unique service ID (SID, service ID), representing an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

[0106] The computing power service is a unit in the computing power node service registry and has a unique service ID. In the computing power service, the service ID is used to identify a specific application service provided by multiple computing power nodes, and at the same time, the access to the computing power service is also initiated through the service ID. When the computing power of the local node cannot meet the computing power request, the computing power requirements corresponding to the computing tasks carried in the computing power request can be packaged into the computing power service. The service ID is a multicast address, and requests for a certain service ID can be responded to by different computing power nodes, that is, it is impossible to predict the computing power nodes providing the computing power service. The process of determining which computing power node to respond to the computing power request by sensing the status of network resources and computing resources is called computing power service scheduling. During the computing power service scheduling, the most suitable computing power node will be selected, that is, the computing power node with the highest scheduling priority will be selected. At the same time, the computing power service also needs to listen to and process computing power requests from other computing power nodes. The computing power service binds the local IP address until it obtains relevant information about the service ID requested by other computing power nodes and accesses a specific computing power node providing the target computing power service. During the service processing, the network resources and computing resources of the local node are updated and synchronized to the computing power network to avoid untimely computing power resource information.

[0107] Since the service ID is a multicast address, the computing power entry needs to jointly consider the computing load and network cost to determine which computing power exit of the service ID to forward the computing power request to.

[0108] A preferred embodiment of a computing power perception system for routers, gateways, IP cameras, and ONUs according to the present invention includes the following modules:

[0109] The computing power node networking module is used to form a computing power network based on the SDN protocol after the computing power nodes of physical devices or virtual devices are started. The physical devices are routers, gateways, IP cameras, or ONUs.

[0110] The network resource detection module is used to start a network detection thread for each computing power node in the computing power network, detect the network resources through the network detection thread, and store the network resources.

[0111] The computing resource calculation module is used to calculate the computing resources of each local node in real time and store the computing resources.

[0112] The service status information table creation module is used to create a service status information table based on the network resources and computing resources for each computing power node to perform computing power perception.

[0113] A computing power evaluation module, which is used to evaluate the computing power of a computing power request based on the service status information table of the local node when each of the computing power nodes receives the computing power request, and generate a computing power evaluation result;

[0114] A load balancing module, which is used to calculate the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow viscosity, and dynamic anycast availability for each of the computing power nodes, and forward the computing power request based on the scheduling priorities and the computing power evaluation result;

[0115] A computing power request execution module, which is used for the computing power node to receive the forwarded computing power request through the computing power entrance, call the corresponding computing power service based on the computing power request and execute it, generate an execution result, and output the execution result through the computing power exit.

[0116] The computing power node networking module is specifically used for:

[0117] After a computing power node with a device type of a physical device or a virtual device is started, through a computing power adapter, it obtains a broadcast data packet broadcast by the remaining computing power nodes based on the SDN protocol, establishes a two-way connection with the corresponding computing power node based on the broadcast data packet, and exchanges the node information of the computing power node, thereby forming a computing power network; the physical device is a router, a gateway, an IPC, or an ONU; the virtual device is deployed on a server; the broadcast data packet carries at least a preamble, an access address, a protocol data unit, a CRC check code, a device name, a service UUID, and a flag bit;

[0118] Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

[0119] The computing power node is a basic functional entity in the computing power network, usually integrated in an edge computing node, can provide the load condition of the computing resources of the local node in real time, and can also provide computing power services to other computing power nodes; when there is a computing power request in the entire computing power network, it can be scheduled and allocated through the computing power resource pool (edge computing power resource pool / cloud computing power resource pool).

[0120] The Computing Power Adapter is an important functional entity in the computing power network, mainly used to help terminal devices (computing power nodes) access the computing power network and achieve efficient scheduling and management of computing power resources. The computing power adapter helps terminal devices access the computing power network by maintaining bound IP (BIP) information, identifying initial request packets, etc. The computing power adapter can be implemented as part of the computing power network (internal mode) or deployed on a separate device (external mode), such as a user plane function (UPF) in a mobile network or a broadband remote access server (BRAS) in a fixed network. In the external mode, the computing power adapter can be placed closer to the terminal device, while the computing power nodes can be placed at an aggregation point and connected to multiple computing power adapters; compared with the internal mode, the external computing power adapter can retain less client binding information, and the functional processing is relatively independent with weak coupling.

[0121] The network resource detection module is specifically used for:

[0122] When each computing power node in the computing power network starts up, a network detection thread is synchronously started respectively. Through the network detection thread, network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput, and jitter are detected in real time, and the network resources are stored.

[0123] Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay. Specifically, in implementation, it can be based on average calculation or norm calculation to improve the reliability of jitter calculation; the two-dimensional array of the transmission delay records the node end point (the computing power node at the transmission end) and the round-trip response time, and the transmission delay within a certain period of time is stored by accumulating the two-dimensional array per second; the network bandwidth can be detected by downloading model data, that is, by downloading model data with a large amount of data for speed measurement.

[0124] The computing resource calculation module is specifically used for:

[0125] Each computing power node calculates in real time the computing resources of this node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate, and load information, and stores the computing resources.

[0126] The load information includes at least the number of sessions being served, the number of queries per second, and the computing delay.

[0127] The service status information table creation module is specifically used for:

[0128] Each computing power node creates a service status information table containing a routing sub-table based on the network resources and computing resources for computing power perception, and shares the service status information table to the computing power network through the SDN protocol;

[0129] The service status information table is updated and synchronized regularly. To avoid the impact of network fluctuations, the update and synchronization of the service status information table of the computing power node are only performed when the index change exceeds the threshold or the update timer expires.

[0130] The computing power evaluation module is specifically used for:

[0131] When each computing power node receives a computing power request carrying at least a computing task, a computing time limit, and a computing priority, it performs a computing power evaluation on the computing power request based on the service status information table of the node itself, and determines whether the network resources and computing resources of the node itself can meet the computing task carried by the computing power request, and then generates a computing power evaluation result of sufficient computing power or insufficient computing power.

[0132] The load balancing module is specifically used for:

[0133] Each computing power node calculates the weights of the network bandwidth, network rate, data flow viscosity, and dynamic anycast availability of the remaining computing power nodes respectively, performs a weighted sum on the network bandwidth, network rate, data flow viscosity, and dynamic anycast availability based on the weights, and then calculates the scheduling priorities of the remaining computing power nodes;

[0134] The specific process of calculating the weight of the network bandwidth is as follows: for network types such as 5G networks and Wi-Fi networks, the computing power node jumps to the transmission delay. Assume that the transmission delay from point A to point B is 5ms, and the data storage is AB5. After merging the vertices, the transmission delay between any two connectable computing power nodes is obtained through the shortest path algorithm. As long as it is judged that the transmission delay of the same vertices in the database is relatively large, it can be replaced; in the initial state, the states of all computing power nodes are the same, and only the transmission delay is used to increase the weight. In short, there are multiple paths connecting point A and point B, and the corresponding weights are set based on the transmission delays of each path.

[0135] The specific process of calculating the weight of the network rate is as follows: The same template data is sent to two adjacent computing power nodes. The computing power node with slow efficiency is set as the base value 1, and the other computing power node is given a relative value (less than 1). When requesting the template data, the computing power required for the request response is selected, and the relevant template data is pushed, and so on. After pairwise comparison, the relative value of the computing power delay is obtained, and the weight of the network rate is set based on this relative value.

[0136] In specific implementation, considering the routing selection mechanism of the optimal computing power node, the routing sub-table of the computing power node contains the status information data of computing resources and network resources, including the evaluation parameters of computing resource performance (CPU, GPU, NPU, memory, etc.), the remaining computing capacity (if the computing power has multi-core and multi-thread processing capabilities, this item is enabled), the computing latency (network rate), etc. By performing a weighted comprehensive calculation of network resources and computing resources, the optimal computing power node is selected for routing forwarding and task execution. When the routing node receives the data packet of the computing task, it parses the task TaskID, computing task type, data flow stickiness requirement attribute, computing power requirement range, longest computing duration, end notification type, etc. of the data packet, and based on the pre-obtained correspondence between the computing task type and the computing performance of other computing power nodes, determines at least one other computing power node and its computing performance corresponding to the computing task type. Based on the computing performance of other nodes and the network link status between this node and other computing power nodes, the target computing power node is determined after comprehensive consideration; the address of the target computing power node is the routing destination address of the data packet, and then the data packet is forwarded based on the target address.

[0137] The data flow stickiness and dynamic anycast availability are used to accelerate the efficiency of selection and scheduling. The dynamic anycast availability can be evaluated based on latency, bandwidth, packet rate, and packet loss rate.

[0138] Data flow stickiness refers to the business data relationship that needs to be called long-term and frequently. For the business with high data flow stickiness, high frequency, and small granularity and the computing power, the weight can be strengthened to avoid problems such as stream interruption, packet loss, and traffic chaos. Therefore, the service can submit the data flow stickiness requirements to the computing power node, including the stream stickiness type, timeout time, frequency, etc., generate and write them into the data flow stickiness requirement data table and store and copy the replicas at the computing power edge perception layer.

[0139] Dynamic anycast realizes routing computing power requests to equivalent service instances (computing power services) by dynamically calculating network resources and computing resources. Traditional anycast is usually used for single-request, single-response communication. When the network state changes, the relative positions of computing power nodes change; while in a computing power-aware network, multi-request, multi-response communication is required between terminal devices and computing power nodes; therefore, in the data plane of a computing power-aware network, it is necessary to maintain data flow stickiness to ensure that computing power requests are processed by the same computing power node, and this computing power node is determined when the first anycast request is received at the entrance of the sensing layer; after listening to the connection, the service access of different terminal devices with the same service ID at the entrance of the sensing layer can be scheduled to different computing power exits, and this feature is called dynamic anycast. Dynamic anycast receives requests asynchronously in multiple threads. According to the requirements of edge computing power, the number of thread pools is configured to avoid a large number of requests affecting the use of memory space. The anycast service table contains flows, service IDs, computing power entrances, computing power exits, etc., and is stored on the computing power service or computing power adapter in the sensing layer, usually less than 100. That is, dynamic anycast dynamically adjusts the optimal path within the anycast group by real-time monitoring of network status (such as latency, load, etc.); for example, the anycast protocol based on mobile agents reduces the resource consumption of global information synchronization through local information collection and dynamic update of the routing table1; in a wireless sensor network, the area-disjoint strategy combines the neighbor node influence factor to optimize the anycast path selection and reduce the transmission collision caused by hidden terminals.

[0140] When the computing power evaluation result is sufficient computing power, call the corresponding computing power service through the computing power entrance of this node to execute the computing power request, generate an execution result, and output the execution result through the computing power exit of this node; when the computing power evaluation result is insufficient computing power, forward the computing power request based on the scheduling priority;

[0141] Specifically in implementation, the computing power exit selected by the computing power entrance is not necessarily the one with the lowest load; the computing power request can also be sent to a computing power exit with relatively lower computing load to avoid network fluctuations.

[0142] In the computing power request execution module, both the computing power entrance and the computing power exit are API interfaces, and the call status of the computing power entrance and the computing power exit is fed back through the status callback interface;

[0143] The computing power entrance is used for real-time addressing and computing power scheduling of the computing power service corresponding to the computing power request; the computing power exit is used for querying, aggregating, outputting, and forwarding the execution result across the network;

[0144] Each of the computing power services has a unique service ID (SID, service ID), representing an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

[0145] The computing power service is a unit in the computing power node service registry and has a unique service ID. In the computing power service, the service ID is used to identify a specific application service provided by multiple computing power nodes, and at the same time, the access to the computing power service is initiated through the service ID. When the computing power of this node cannot meet the computing power request, the computing power demand corresponding to the computing task carried in the computing power request can be packaged into the computing power service; the service ID is a multicast address, and the request for a certain service ID can be responded to by different computing power nodes, that is, it is impossible to predict the computing power nodes providing the computing power service. The process of determining which computing power node to select to respond to the computing power request by sensing the status of network resources and computing resources is called computing power service scheduling. During the computing power service scheduling, the most suitable computing power node will be selected, that is, the computing power node with the highest scheduling priority will be selected. At the same time, the computing power service also needs to listen to and process the computing power requests from other computing power nodes; the computing power service binds the local IP address until it obtains the relevant information of the service ID requested by other computing power nodes and accesses the specific computing power node providing the target computing power service. During the service processing, the network resources and computing resources of this node are updated and synchronized to the computing power network to avoid untimely computing power resource information.

[0146] Since the service ID is a multicast address, the computing power entrance needs to jointly consider the computing load and network cost to determine which computing power exit of the service ID to forward the computing power request to.

[0147] In summary, the advantages of the present invention are as follows:

[0148] 1. After the computing power nodes are started, a computing power network is formed through networking based on the SDN protocol. Each computing power node in the computing power network starts a network detection thread respectively, detects and stores network resources through the network detection thread, and each of the computing power nodes calculates and stores its own computing resources in real time. A service status information table is created based on the network resources and computing resources for computing power perception. Then, when each computing power node receives a computing power request, it performs computing power evaluation on the computing power request based on the service status information table of its own node to generate a computing power evaluation result. Then, each computing power node calculates the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data stream viscosity, and dynamic anycast availability, and forwards the computing power request based on the scheduling priorities and the computing power evaluation result. Finally, the computing power node receives the forwarded computing power request through the computing power entrance, calls the corresponding computing power service based on the computing power request and executes it, and outputs the execution result through the computing power exit. That is, based on the SDN protocol, the distributed computing power nodes with different architectures and different algorithms are connected, the node information is aggregated, synchronized, and heterogeneousized, and the network resources and computing resources are dynamically and real-time perceived to create a service status information table, and then the computing tasks carried by the computing power requests are overall allocated and scheduled, forming a computing power network in which the computing power resources (network resources and computing resources) can be perceived, allocated, and scheduled. When the computing power of a certain computing power node is insufficient, the computing power of other computing power nodes is borrowed for load balancing; and the computing power scheduling is performed through the service status information table, that is, for specific data serialization, the service status information table can be flexibly defined to meet different expansion requirements, ultimately greatly improving the load balancing ability and scalability of the computing power nodes, thereby greatly improving the computing efficiency of edge computing and greatly reducing the computing cost of edge computing.

[0149] 2. Through a lightweight SDN protocol, network exploration, data encapsulation, data reporting, and data synchronization are carried out, the status of network resources and computing resources is monitored and perceived in real time, and the proximal edge layer in the computing power network is reported to push the computing tasks to the network edge of the computing power network, making the computing tasks closer to the data source or users, effectively reducing the transmission delay, and ensuring the quality of service (QoS) and user experience (QoE). Especially in real-time and bandwidth-sensitive applications, the purpose of computing power load balancing is achieved, and a certain scale of computing power efficiency breakthrough is realized.

[0150] 3. Build a computing power network through the SDN (software-defined networking) protocol. Since the SDN protocol can program the control plane, the separation of the data plane and the control plane will help abstract the underlying network infrastructure resources and provide APIs in the form of programmable network objects. By programming network objects, there is no need to focus on how to implement the specific logic of underlying network devices such as routers and switches, and it has better flexibility and controllability. Moreover, the SDN protocol design can focus on the content and forwarding logic of data without considering the underlying decision-making strategies. All decision-making strategies are determined and processed by the SDN controller on the control plane, which makes the configuration and management of the computing power network very agile and efficient.

[0151] Although the specific implementation manners of the present invention have been described above, those skilled in the art of this technology should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.

Claims

1. A computing power perception method for routers, gateways, IP cameras, and ONUs, characterized in that: It includes the following steps: Step S1: After a computing power node with a device type of physical device or virtual device is started, it forms a computing power network based on the SDN protocol; the physical device is a router, gateway, IPC or ONU; Step S2: Each computing power node in the computing power network starts a network detection thread respectively, detects network resources through the network detection thread, and stores the network resources; Step S3: Each computing power node calculates the computing resources of its own node in real time and stores the computing resources; Step S4: Each computing power node creates a service status information table based on the network resources and computing resources to perform computing power perception; Step S5: When each computing power node receives a computing power request, it performs computing power evaluation on the computing power request based on the service status information table of its own node to generate a computing power evaluation result; Step S6: Each computing power node calculates the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data flow stickiness and dynamic anycast availability, and forwards the computing power request based on the scheduling priorities and the computing power evaluation result; Step S7: The computing power node receives the forwarded computing power request through the computing power entrance, invokes the corresponding computing power service based on the computing power request and executes it to generate an execution result, and outputs the execution result through the computing power exit.

2. The computing power awareness method for routers, gateways, IP cameras, and ONUs according to claim 1, wherein: The specific content of step S1 is as follows: After a computing power node with a device type of physical device or virtual device is started, through a computing power adapter, it obtains the broadcast data packets broadcast by the remaining computing power nodes based on the SDN protocol, establishes a two-way connection with the corresponding computing power node based on the broadcast data packets, and exchanges the node information of the computing power node, thereby forming a computing power network; the physical device is a router, gateway, IPC or ONU; the virtual device is deployed on a server; Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

3. The computing power perception method for a router, gateway, IPC, and ONU according to claim 1, wherein: The specific content of step S2 is as follows: When each computing power node in the computing power network is started, it synchronously starts a network detection thread respectively, and detects network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput and jitter in real time through the network detection thread, and stores the network resources; Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay; The specific content of step S3 is as follows: Each computing power node calculates the computing resources of its own node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate and load information in real time, and stores the computing resources; The load information includes at least the number of sessions being served, the number of queries per second and the computing delay.

4. The computing power awareness method for a router, gateway, IPC, and ONU according to claim 1, characterized in that: The specific content of step S4 is as follows: Each computing power node creates a service status information table containing a routing sub-table based on the network resources and computing resources to perform computing power perception, and shares the service status information table to the computing power network through the SDN protocol; The specific content of step S5 is as follows: When each computing power node receives a computing power request carrying at least a computing task, a computing time limit, and a computing priority, it performs a computing power evaluation on the computing power request based on the service status information table of this node, and determines whether the network resources and computing resources of this node can meet the computing task carried by the computing power request, and then generates a computing power evaluation result of sufficient computing power or insufficient computing power.

5. The computing power perception method for a router, gateway, IPC, and ONU according to claim 1, characterized in that: The specific steps of step S6 are as follows: Each of the computing power nodes calculates the weights of the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability of the remaining computing power nodes respectively, performs a weighted sum on the network bandwidth, network rate, data stream viscosity, and dynamic anycast availability based on the weights, and then calculates the scheduling priorities of the remaining computing power nodes. When the computing power evaluation result is sufficient computing power, the corresponding computing power service is called through the computing power entrance of this node to execute the computing power request, and an execution result is generated, and the execution result is output through the computing power exit of this node; when the computing power evaluation result is insufficient computing power, the computing power request is forwarded based on the scheduling priority. In step S7, both the computing power entrance and the computing power exit are API interfaces, and the call status of the computing power entrance and the computing power exit is fed back through the status callback interface. The computing power entrance is used for real-time addressing and computing power scheduling of the computing power service corresponding to the computing power request; the computing power exit is used for processing query, aggregation, output, and full-network forwarding of the execution result. Each of the computing power services has a unique service ID, which represents an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

6. A computing power perception system for routers, gateways, IP cameras, and ONUs, characterized in that: It includes the following modules: The computing power node networking module is used for, after the computing power nodes of physical devices or virtual devices are started, networking based on the SDN protocol to form a computing power network; the physical devices are routers, gateways, IPCs, or ONUs. The network resource detection module is used for each computing power node in the computing power network to start a network detection thread respectively, detect network resources through the network detection thread, and store the network resources. The computing resource calculation module is used for each computing power node to calculate the computing resources of this node in real time and store the computing resources. The service status information table creation module is used for each computing power node to create a service status information table based on the network resources and computing resources for computing power perception. The computing power evaluation module is used for each computing power node to perform a computing power evaluation on the computing power request based on the service status information table of this node when receiving the computing power request, and generate a computing power evaluation result. The load balancing module is used for each computing power node to calculate the scheduling priorities of the remaining computing power nodes based on network bandwidth, network rate, data stream viscosity, and dynamic anycast availability, and forward the computing power request based on the scheduling priority and the computing power evaluation result. The computing power request execution module is used for the computing power node to receive the forwarded computing power request through the computing power entrance, call the corresponding computing power service based on the computing power request and execute it, generate an execution result, and output the execution result through the computing power exit.

7. The computing power perception system for routers, gateways, IP cameras, and ONUs according to claim 6, wherein: The computing power node networking module is specifically used for: After a computing power node with a device type of physical device or virtual device is started, through a computing power adapter, based on the SDN protocol, it obtains broadcast data packets broadcast by other computing power nodes, establishes a two-way connection with the corresponding computing power nodes based on the broadcast data packets, and exchanges node information of the computing power nodes, thereby forming a computing power network; the physical device is a router, gateway, IPC or ONU; the virtual device is deployed on a server; Each of the physical devices forms an edge computing power resource pool, and each of the virtual devices forms a cloud computing power resource pool.

8. The computing power perception system for routers, gateways, IP cameras, and ONUs according to claim 6, characterized in that: The network resource detection module is specifically used for: When each computing power node in the computing power network is started, a network detection thread is synchronously started respectively, and the network resources including at least network bandwidth, bandwidth utilization rate, transmission delay, throughput and jitter are detected in real time through the network detection thread, and the network resources are stored; Both the network bandwidth and the transmission delay are stored based on a two-dimensional array; the jitter is calculated based on the transmission delay; The computing resource calculation module is specifically used for: Each computing power node calculates in real time the computing resources of this node including at least CPU utilization rate, GPU utilization rate, NPU utilization rate, memory occupancy rate and load information, and stores the computing resources; The load information includes at least the number of sessions being served, the number of queries per second and the computing delay.

9. The computing power perception system for a router, gateway, IPC, and ONU according to claim 6, characterized in that: The service status information table creation module is specifically used for: Each computing power node creates a service status information table including a routing sub-table based on the network resources and computing resources for computing power perception, and shares the service status information table to the computing power network through the SDN protocol; The computing power evaluation module is specifically used for: When each computing power node receives a computing power request carrying at least a computing task, a computing time limit and a computing priority, it performs a computing power evaluation on the computing power request based on the service status information table of this node, and judges whether the network resources and computing resources of this node can meet the computing task carried by the computing power request, thereby generating a computing power evaluation result of sufficient computing power or insufficient computing power.

10. A computing power perception system for routers, gateways, IP cameras, and ONUs according to claim 6, characterized in that: The load balancing module is specifically used for: Each computing power node calculates the weights of the network bandwidth, network rate, data stream stickiness and dynamic anycast availability of the other computing power nodes respectively, performs a weighted sum on the network bandwidth, network rate, data stream stickiness and dynamic anycast availability based on the weights, and thereby calculates the scheduling priorities of the other computing power nodes; When the computing power evaluation result is sufficient computing power, call the corresponding computing power service through the computing power entry of this node to execute the computing power request, generate an execution result, and output the execution result through the computing power exit of this node; when the computing power evaluation result is insufficient computing power, forward the computing power request based on the scheduling priority; In the computing power request execution module, both the computing power entry and the computing power exit are API interfaces, and the call status of the computing power entry and the computing power exit is fed back through a status callback interface; The computing power entry is used to perform real-time addressing and computing power scheduling for the computing power service corresponding to the computing power request; the computing power exit is used to process the query, aggregation, output, and full-network forwarding of the execution result; Each of the computing power services is provided with a unique service ID, which represents an application service, and the service ID is bound to the IP address of this node; the service ID is a multicast address.

Images