Calculation power network service anycast method and system

Abstract

The invention provides a computing power network service anycast method and system. The method comprises the steps that a gateway forwards a service request to a controller; the controller calculates the transmission cost and the propagation cost from the gateway to each cloud node and each edge node based on the transmission distance and the bandwidth between the gateway and the cloud nodes and the edge nodes, and calculates a network transmission cost matrix based on the transmission cost and the propagation cost; calculating an arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node; calculating a target anycast forwarding weight strategy matrix based on the arrival rate matrix and the network transmission cost matrix by adopting a Sinkhorn algorithm, wherein the target anycast forwarding weight strategy matrix comprises weight values of each cloud node and each edge node; and determining the number of positions of the cloud nodes and the edge nodes in a queue based on the weight values of the cloud nodes and the edge nodes, constructing the queue, and issuing a processing task corresponding to the service request to the corresponding cloud nodes or the edge nodes based on the queue.

Description

Technical Field

[0001] The present invention relates to the field of computing power network technology, and in particular to a computing power network service anycast method and system. Background Art

[0002] In the information age, globally interconnected mobile computing has penetrated deeply into every aspect of human life. At the same time, the computing model is also changing, from Internet-centric cloud computing, to edge computing where services can be closed nearby and agile and intelligent, to the future linkage between cloud-edge computing capabilities and the end side, the computing model is developing towards a ubiquitous architecture with multi-level deployment of cloud-edge-end. This requires the expansion of end-to-end communications to enhance high-performance computing applications (e.g., emergency smart applications), machines and mobile objects (e.g., self-driving cars and drones), content distribution and streaming services, and focuses on increasing the ability to virtualize and replicate any service to improve efficiency and redundancy.

[0003] The current network routing and forwarding system designed for the Internet is mainly designed for end-to-end communication between single-destination nodes. Network devices route packets carrying application layer requests to a unique peer network node based on their destination addresses. Although load balancers placed at, for example, public cloud entrances can be used to balance the load between homogeneous service instances, this system limits the potential for coordinated collaborative services between cloud-edge, edge-edge, and multi-cloud. Service Anycast allows application service replicas at different locations to set a shared service identifier and forward service requests to one of multiple replicas based on certain policies, which provides users with enhanced quality of experience and service providers with multiple advantages (e.g., load balancing between replicas, robustness against DDoS attacks). However, how to coordinate the differences in service replica instances and design service forwarding mechanisms and forwarding decision calculation methods based on the full decoupling of services and address locations to improve overall service performance has become a key challenge. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a computing power network service anycast method and system to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a computing network service anycast method, the method is based on a pre-built computing network, the computing network includes a gateway, a controller, a plurality of cloud nodes and a plurality of edge nodes, the method comprises the following steps:

[0006] Based on the gateway receiving the service request, the gateway forwards the service request to the controller;

[0007] The controller obtains the working status of each cloud node and edge node in real time, and constructs the cloud nodes and edge nodes into a node set;

[0008] Based on the transmission distance and bandwidth between the gateway and the cloud nodes and edge nodes, the transmission cost and propagation cost from the gateway to each cloud node and edge node are calculated, and based on the transmission cost and propagation cost, a network transmission cost matrix is ​​calculated;

[0009] Calculate the arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node;

[0010] Calculate a target anycast forwarding weight strategy matrix using a Sinkhorn algorithm based on the arrival rate matrix and the network transmission cost matrix, wherein the target anycast forwarding weight strategy matrix includes a weight value of each cloud node and edge node;

[0011] The number of positions of the cloud node and the edge node in the queue is determined based on the weight values ​​of the cloud node and the edge node, and a queue is constructed. Based on the queue, the processing task corresponding to the service request is sent to the corresponding cloud node or edge node.

[0012] By adopting the above scheme, this scheme realizes the organic combination of centralized control and distributed operation through the separation of the software defined network (SDN) control plane and the data plane; in the complex environment of cross-layer nodes, the computing network service anycast forwarding architecture makes full use of the global optimization capability of centralized control and the flexibility of the queue forwarding mechanism based on weighted strategy in the distributed data plane, so that the system can perform fine-grained balanced scheduling of burst traffic and traffic fluctuations from the perspective of global optimality, while ensuring stable service quality under high load conditions and improving resource utilization.

[0013] In some embodiments of the present invention, the steps of the method include: when the controller receives a service request for the first time, the controller parses the service type corresponding to the service request, calculates a network transmission cost matrix and an arrival rate matrix corresponding to the service type, and calculates a queue corresponding to the service type;

[0014] When the controller receives a service request again, it analyzes the service type corresponding to the service request and determines whether there is a queue for the service type currently;

[0015] If it exists, the processing task corresponding to the service request is directly sent to the corresponding cloud node or edge node based on the existing queue;

[0016] If it does not exist, construct a queue corresponding to the service type.

[0017] In some embodiments of the present invention, in the step of sending the processing task corresponding to the service request to the corresponding cloud node or edge node based on the queue, the cloud node or edge node at the head of the queue is used as the cloud node or edge node to be sent, and the cloud node or edge node is moved to the tail of the queue to update the queue.

[0018] In some embodiments of the present invention, in the step of calculating the transmission cost and propagation cost from the gateway to each cloud node and edge node based on the transmission distance and bandwidth between the gateway and the cloud node and the edge node, and calculating the network transmission cost matrix based on the transmission cost and propagation cost:

[0019] Based on the input data volume and output data volume transmitted by the gateway to each cloud node and edge node based on the preset data packet, the transmission cost is calculated based on the input data volume and output data volume and the bandwidth; the time consumed by the preset data packet from the gateway to each cloud node and edge node is used as the transmission cost;

[0020] The gateway transmits a cost value to each cloud node and edge node based on the transmission cost and the propagation cost, and constructs the cost matrix of network transmission.

[0021] In some embodiments of the present invention, in the step of calculating the transmission cost based on the input data volume and the output data volume transmitted by the gateway to each cloud node and edge node based on the preset data packet, the transmission cost is calculated based on the following formula:

[0022]

[0023] in, represents the transmission cost value from the gateway router i to the cloud node and edge node j; represents the preset input data volume of data packets for service k transmitted by gateway router i to cloud node and edge node j; represents the preset output data volume of data packets for service k transmitted from gateway router i to cloud node or edge node j; It represents the bandwidth of the preset data packet transmitted from the gateway router i to the cloud node or edge node j.

[0024] In some embodiments of the present invention, in the step of calculating the target anycast forwarding weight strategy matrix based on the arrival rate matrix and the network transmission cost matrix using the Sinkhorn algorithm, the arrival rate matrix and the network transmission cost matrix are used as two diagonal matrices of the Sinkhorn algorithm, and an objective function is constructed based on the arrival rate matrix and the network transmission cost matrix, and the final objective function value is iterated based on the Sinkhorn algorithm to obtain the target anycast forwarding weight strategy matrix.

[0025] In some embodiments of the present invention, in the step of constructing an objective function based on the arrival rate matrix and the network transmission cost matrix, the objective function is constructed as:

[0026] L=MIN <P k , C k >

[0027] Among them, L represents the objective function value, P k represents the arrival rate matrix for service k, C k represents the network transmission cost matrix for service k, <P k , C k > means calculating the inner product of the matrix, and MIN means minimization.

[0028] In some embodiments of the present invention, in the step of calculating the arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node, the arrival rate matrix is ​​calculated based on the following formula:

[0029] P k =α k ·X k ;

[0030] Among them, P k represents the arrival rate matrix, α k represents the arrival rate vector, X k Represents the forwarding strategy matrix.

[0031] In some embodiments of the present invention, in the step of calculating the network transmission cost matrix based on the transmission cost and the propagation cost, the transmission cost value transmitted by the gateway's router to each cloud node and edge node is calculated, and the propagation cost value transmitted by the gateway's router to each cloud node and edge node is measured through a preset data packet for service k. For each cloud node or edge node, the sum of the transmission cost value and the propagation cost value is calculated as the value of each position in the network transmission cost matrix to construct the network transmission cost matrix.

[0032] The second aspect of the present invention also provides a computing power network service anycast system, which includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0033] The third aspect of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps implemented by the aforementioned computing power network service anycast method.

[0034] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention can be specifically pointed out and obtained in the specification and the accompanying drawings.

[0035] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0037] Figure 1 This is a schematic diagram of an implementation method of the computing power network service anycast method of this solution;

[0038] Figure 2 This is a schematic diagram of the architecture of the computing network service anycast method of this solution;

[0039] Figure 3 This is a schematic diagram of the update process of the computing network service anycast forwarding queue of this solution;

[0040] Figure 4 This is a schematic diagram of the Packet_in message processing flow of the computing network service control plane of this solution;

[0041] Figure 5 Schematic diagram of the service anycast strategy calculation algorithm based on optimal transmission in this scheme. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0043] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0044] Introduction to existing technology:

[0045] Existing plan 1:

[0046] The first existing solution involves the field of computing power scheduling. In order to solve the problem that static rule routing cannot adapt to dynamic environmental changes, the computing power routing method includes obtaining first computing power resource information and task characteristic information of the computing task to be processed; obtaining the probability distribution of forwarding the computing task to be processed to each neighboring node according to the task characteristic information and the first computing power resource information; sampling in the probability distribution to obtain the target forwarding node, and forwarding the task characteristic information to the target forwarding node. The present invention can adapt to dynamic environmental changes in the computing power scheduling system, improve the resource utilization and security of the computing power scheduling system, and optimize the performance of the computing power scheduling system.

[0047] Existing solution 2:

[0048] The existing scheme 2 aims to achieve efficient and reliable computing resource allocation and task scheduling through the learning and decision-making of intelligent agents. The method includes: perceiving computing network resources; selecting computing nodes according to task requirements and node evaluation information; making routing path decisions using a controller based on a graph neural network containing a deep reinforcement learning agent to determine the routing path from the start node to the destination node; forwarding the determined routing path to the nodes in the computing network, and updating the state information of the nodes and the network. The present invention is based on a reinforcement learning algorithm, and defines states, actions, and reward functions through a graph neural network, and can learn the optimal routing strategy from the state space. Experimental verification has shown that the present invention can effectively complete the allocation of more computing tasks, and maintain stability and reliability under different network topologies and task load conditions.

[0049] Technical disadvantages of existing solution 1:

[0050] (1) Lack of dynamic load balancing capabilities and poor policy stability: This computing power routing method only allocates tasks based on node probability distribution sampling, and fails to fully consider the real-time load information of service instances. This approach may cause some nodes to continue to receive requests under high load, reducing overall resource utilization efficiency and easily causing local congestion.

[0051] (2) Lack of a mechanism to ensure service quality: This method mainly relies on static computing resource information and task feature information in forwarding decisions, and lacks detailed quality of service (QoS) control measures. It is impossible to strike a balance between service quality and resource allocation, making it difficult to meet service requirements in multi-level and complex application scenarios.

[0052] (3) Unable to adapt to the service coordination requirements across multiple layers of nodes: This method mainly focuses on the scheduling optimization of a single computing resource, lacks the ability to coordinate and process edge, cloud and multi-level resources, and is difficult to adapt to the effective collaboration of computing resources across cloud-edge, edge and multi-cloud environments.

[0053] Technical disadvantages of the existing solution 2:

[0054] (1) High computational and resource overhead: The introduction of deep reinforcement learning and graph neural networks has greatly increased computational complexity and resource overhead, especially in multi-layer, large-scale network topologies. This high overhead may affect the real-time performance of the system and is not conducive to quickly responding to task requirements in a dynamic computing environment.

[0055] (2) Model training relies on a large amount of data: The training of reinforcement learning models requires a large amount of task allocation and routing decision data to learn effective strategies. In actual network deployment, collecting enough data may cause delays, or insufficient data may affect the generalization and deployment effect of the model.

[0056] (3) Difficulty in coping with rapidly changing network conditions: This method relies on the training of deep reinforcement learning models, so the model update speed is usually slow and it is difficult to adapt to rapidly changing network conditions and task requirements in a timely manner, resulting in low adaptability of real-time routing decisions.

[0057] (4) Low transparency in path selection: Due to the “black box” nature of reinforcement learning models and graph neural networks, it is difficult to explain their specific decision paths and selection mechanisms. This lack of transparency may affect the management and maintenance of service operators during the deployment process and increase the complexity of routing troubleshooting.

[0058] (5) Limited support for anycast across multiple layers of nodes: This method focuses on routing path selection and resource allocation for tasks, and lacks support for multi-layer coordination and anycast forwarding of service instances in the computing network, which may make it difficult to meet the load balancing requirements of different service instances.

[0059] In summary, existing technologies lack the ability to dynamically and finely schedule service instances: as computing resources are gradually embedded in network infrastructure, users increasingly want efficient, low-latency services, rather than just focusing on specific physical locations. Traditional network architectures are mainly centered around end-to-end communication, and are difficult to adapt to the dynamic scheduling needs of different service instances. Since the location, load conditions, and user requests of service instances often change dynamically, current load balancers and static routing mechanisms are difficult to flexibly allocate resources, resulting in resource waste and unstable service quality. In addition, the current design of anycast forwarding of computing power services (some work refers to computing power routing) often uses a single-purpose service instance to process service requests on computing power network access routers within scheduling time slots, making it difficult to perform fine-grained balancing of differentiated service instance processing capabilities.

[0060] Existing technologies lack resource coordination across multiple layers of nodes: Computing networks usually include multiple layers of resource nodes, such as cloud and edge nodes. Nodes at different levels vary significantly in terms of computing and storage capabilities, geographical distribution, etc. However, the current network architecture lacks a collaborative scheduling mechanism across multiple layers of nodes, making it difficult for edge node computing resources to fully collaborate with cloud nodes. The lack of such a collaborative mechanism not only increases the burden on individual nodes, but also limits the efficiency of global resource utilization, ultimately reducing the overall quality of service.

[0061] The instability of the selection strategy caused by the mismatch between the dimensions of the network and computing load indicators in the existing technology: Although the introduction of a dynamic selection strategy based on real-time load and network conditions can optimize the service response effect, the indirect processing time (such as CPU utilization) indicators used in the load evaluation of computing resources do not fully match the actual network and computing conditions. In addition, due to the delay in updating computing load information, the selection strategy in such systems is often too greedy and unstable, resulting in poor adaptation of computing resources and networks, and thus impairing the stability of services and user experience.

[0062] like Figure 1 and 2 As shown, the present invention proposes a computing network service anycast method, which is based on a pre-built computing network, wherein the computing network includes a gateway, a controller, a plurality of cloud nodes and a plurality of edge nodes. The steps of the method include:

[0063] In the specific implementation process, there are multiple service replica instances for the same service deployed concurrently in the computing network, and different service instances can be deployed simultaneously on different computing nodes, such as edge nodes (EN) and cloud nodes (CN), and have different request processing capabilities based on the computing resources and network resources occupied by the instance.

[0064] Step S100, based on the gateway receiving the service request, the gateway forwards the service request to the controller;

[0065] In the specific implementation process, the computing power network (CPN) access router located near the user (such as a mobile edge computing site) has three main functions:

[0066] 1. Request statistics and reporting: The access router gateway can monitor key data such as user access frequency, request types, and request processing time in real time by collecting statistics of user service requests. These statistics are regularly reported to the centralized controller so that the controller can make global optimization decisions, such as load balancing and service migration. In addition, the access router can implement local caching according to different service request flows to improve access performance.

[0067] 2. Anycast forwarding queue processing: The access router gateway determines the next hop target of the service request based on the anycast forwarding queue received from the controller. During the forwarding process, it adopts a circular forwarding queue mechanism so that service requests can be flexibly allocated between different computing nodes, thereby improving the concurrency and response speed of service processing.

[0068] 3. Request encapsulation and forwarding: In some cases, the access router gateway needs to encapsulate the original service request, especially when forwarding requests between different network domains. After encapsulating the outer header, the access router forwards the request to the egress router to ensure that the request can be effectively delivered across different network environments.

[0069] The computing network egress router deployed near the computing nodes has two main functions:

[0070] 1. Unpacking and forwarding: The main responsibility of the egress router is to unpack the encapsulated request from the access router and restore the original request information. The unpacked request is forwarded to the original request access router according to the service type and topology to ensure that the service instance can correctly receive and process the request.

[0071] 2. Service information reporting: In addition to forwarding, the egress router also needs to collect service information deployed on the computing node, such as service type, processing capacity, service instance load, etc., and report this information to the controller. This data helps the controller to dynamically schedule resources and manage service instances to achieve global optimal resource allocation.

[0072] Step S200, the controller obtains the working status of each cloud node and edge node in real time, and constructs the cloud nodes and edge nodes into a node set;

[0073] In the specific implementation process, this solution uses OpenFlow and Ryu controller to implement the computing network service anycast forwarding system, and combines key features such as CPN topology awareness, service ID storage and matching. The Locust stress test tool is used to evaluate the service processing capacity and assign an anycast IP to the service ID.

[0074] Step S300, based on the transmission distance and bandwidth between the gateway and the cloud node and the edge node, the transmission cost and propagation cost from the gateway to each cloud node and the edge node are calculated, and the network transmission cost matrix is ​​calculated based on the transmission cost and propagation cost;

[0075] Step S400, calculating an arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node;

[0076] Step S500, calculating a target anycast forwarding weight strategy matrix using a Sinkhorn algorithm based on the arrival rate matrix and the network transmission cost matrix, wherein the target anycast forwarding weight strategy matrix includes a weight value of each cloud node and edge node;

[0077] In the specific implementation process, due to the limited programmability of the data plane of OpenFlow, this solution uses the control plane application written by RYU to record and maintain the service anycast forwarding table of different CPN access routers and perform forwarding queue updates to implement the service anycast mechanism. The service anycast forwarding entry class contains the following entry attributes:

[0078] 1. The software switch data channel ID (dpid) to which this service forwarding entry belongs;

[0079] 2. The service ID to which this service forwards the entry;

[0080] 3. The transport layer protocol type of the service matched by this service forwarding entry;

[0081] 4. An optional list of service instance IDs for services matched by this service forwarding entry;

[0082] 5. The forwarding strategy currently used by the forwarding entries of this service;

[0083] 6. The service forwarding destination ID (newest_target) of the service forwarding entry that was most recently set;

[0084] 7. Indicates whether the current entry needs to be updated (the default is to update);

[0085] 8. Record the software switch flow table matching entry corresponding to this forwarding;

[0086] 9. Record the software switch reverse flow table matching entry corresponding to this forwarding;

[0087] 10. Record the time when this forwarding entry triggers the flow table to be sent based on the Openflow packet_in message;

[0088] 11. Other related settings of the forwarding items of this service;

[0089] 12. The forwarding queue being used by the forwarding entry of this service;

[0090] The key service forwarding queue update generation function of the service anycast forwarding entry class is as follows: Figure 3 As shown. In this function, firstly, the forwarding policy forwarding_policy needs to be input. The forwarding policy element indicates the probability weight of the service request forwarded to the corresponding service instance. Then, a new forwarding queue forwarding_queue is created. The policy is generated by rounding the product of the forwarding policy and the number of instances. The policy element indicates the number of service instances with different destinations in the forwarding queue. When the sum of the elements of the policy is less than a small value, such as 10^(-5), the loop is stopped. Otherwise, the policy elements are traversed, and the policy elements are decremented by 1 and the service instances indicated by the policy elements are added to the temporary queue temp. After the policy elements are traversed, the temp queue is shuffled and added to the forwarding queue forwarding_queue. The above operation ensures that even if the forwarding queue mechanism cannot be implemented on the data plane, the implementation of the mechanism on the control plane can reduce the imbalance caused by the service requests being sent to the same destination instance in the same forwarding decision cycle, thereby achieving a more balanced service forwarding effect.

[0091] In software-defined network programming, when the software-defined CPN access router cannot process a data packet according to the local flow table, it reports it to the controller, and the controller listens to such packet_in messages to complete the implementation of the software-defined network function. Figure 4 As shown, in this solution, when the controller determines that the destination IP address of the ARP request is a service anycast address, it generates an ARP Reply message based on the MAC address corresponding to the service anycast address, and requires the access router to forward the response message. When the controller determines that the message is a TCP / UDP request and the request destination is a service anycast IP, it first records the CPN access router that received the message, creates a service routing entry based on the service subscription retrieval table, and then generates a service forwarding policy based on the forwarding policy calculation algorithm. According to the service forwarding policy, the service anycast forwarding queue update process is executed to generate a service forwarding queue, and the following two computing power service forwarding flow table entries are issued:

[0092] 1. Item 1: Replace the destination address with the service instance address at the head of the forwarding queue, issue this item and request the switch to continue forwarding the data packets received by this access router;

[0093] 2. Item 2: If the destination address of the request matches the host address directly connected to this CPN access router, and the source address matches the service instance address maintained by this access router, then the source address is replaced with the service anycast address and this item is issued.

[0094] Step S600, based on the weight values ​​of the cloud nodes and edge nodes, determine the number of positions of the cloud nodes and edge nodes in the queue, build a queue, and send the processing task corresponding to the service request to the corresponding cloud node or edge node based on the queue.

[0095] In the specific implementation process, this solution is also divided into data plane functions and control plane functions in the computing network service anycast forwarding system architecture. The data plane functions include:

[0096] 1. Service request processing based on circular forwarding queue: In the traditional target metric forwarding mechanism, the selection of the target node is usually based on static metrics such as delay and number of hops. In the computing power network designed by the present invention, a dynamic circular forwarding queue mechanism is introduced in the data plane. When a service request arrives and matches the service ID, the queue will use the target computing node at the front as the forwarding target. This mechanism ensures that under a good request forwarding strategy design, the request can be evenly distributed to multiple service copies in an update granularity method to prevent overload of certain nodes. After the forwarding is completed, the target will be moved to the end of the queue to ensure the persistence of the policy. The service anycast forwarding table is shown in Table 1. The forwarding table is subordinate to the data channel ID of the CPN access router and contains information such as the destination matching service ID, forwarding queue, queue issuance time, queue expiration time, and service ID matching times.

[0097] Table 1

[0098]

[0099] Service-based encapsulation and decapsulation: In some complex network environments, the data plane is also responsible for encapsulating and decapsulating the external headers of cross-domain service requests to ensure that the requests can be processed across different network areas or devices.

[0100] Control plane functions include:

[0101] 1. Information perception layer functions, including:

[0102] a. Service request rate prediction: By analyzing the service request statistics reported by the access router, information perception can predict the request arrival rate in the future in real time. This helps to pre-schedule computing resources in advance, optimize service response speed, and reduce the startup time of service replicas.

[0103] b.CPN topology awareness: The information awareness layer continuously monitors the network topology of the CPN, including the location of computing nodes, the status of network links, and the computing load of each node. This awareness function can provide real-time data support for the calculation of forwarding strategies.

[0104] c. Service instance discovery: The information perception layer uses the reported information from each egress router to understand the current deployed service instances, including the number of service replicas, distribution locations, and processing capabilities. This helps to dynamically adjust the anycast forwarding strategy and achieve load balancing.

[0105] 2. Forwarding strategy calculation layer functions, including:

[0106] a. Service subscription retrieval: When a user request arrives at the access router, if the router does not have a forwarding entry that matches the service request, the request is reported to the control plane. The control plane first retrieves the service subscription to retrieve the appropriate service instance that can match the user request. As shown in Table 2, the computing power network control plane service subscription retrieval table, the control plane function also needs to record the CPN access router ID information in the subscription service router list. The control plane service subscription retrieval table includes the service ID, the optional service instance list, the service instance unit time processing capacity, the information update time, the information expiration time, and the CPN access router list that subscribes to this service.

[0107] Table 2

[0108]

[0109] b. QoS identification and forwarding queue generation: The forwarding policy calculation layer identifies the priority of the request according to the QoS requirements of the service, such as bandwidth, delay, jitter, etc., and generates an anycast forwarding queue based on the forwarding policy calculation method to ensure that the service request can be forwarded to the target computing node in the most balanced way.

[0110] c. Forwarding strategy optimization: The computing layer is also responsible for continuously optimizing the existing forwarding strategies, such as performing global path optimization through AI algorithms to improve the utilization efficiency of network resources.

[0111] By adopting the above scheme, this scheme realizes the organic combination of centralized control and distributed operation through the separation of the software defined network (SDN) control plane and the data plane; in the complex environment of cross-layer nodes, the computing network service anycast forwarding architecture makes full use of the global optimization capability of centralized control and the flexibility of the queue forwarding mechanism based on weighted strategy in the distributed data plane, so that the system can perform fine-grained balanced scheduling of burst traffic and traffic fluctuations from the perspective of global optimality, while ensuring stable service quality under high load conditions and improving resource utilization.

[0112] like Figure 2 As shown in the figure, in the computing network service anycast forwarding system architecture, the service request processing flow is as follows:

[0113] 1. Service request arrives: When a user sends a service request, the request first arrives at the CPN access router closest to the user. The access router first checks its local forwarding table to see if there is a forwarding path for the service. If there is a valid path for the service in the forwarding table, it is forwarded directly according to the forwarding queue.

[0114] 2. Service subscription request: The access router uses a publish-subscribe model to record requests that are currently being processed but not yet completed. This mechanism reduces the search space for service requests in the network, thereby speeding up request processing. Specifically, if the access router does not find a matching service forwarding path, it sends a service subscription request to the controller. The request contains key information such as the type of service required by the user and QoS requirements. The controller retrieves the deployed service instances by subscribing to the service, and selects a group of service instances based on factors such as the geographical location, computing power, and current load of the service, and records them in the controller's service subscription record list.

[0115] 3. Generate and send anycast forwarding queue: After receiving the service subscription request, the controller combines the information of CPN topology perception and service instance discovery module to generate anycast forwarding queue for the user service request. This queue contains multiple target computing nodes, and the appearance frequency of different target computing nodes in the forwarding queue is set according to the forwarding strategy requirements to ensure balanced service quality. The controller then sends this forwarding queue to the access router.

[0116] 4. Forwarding execution: After receiving the anycast forwarding queue returned by the controller, the access router forwards the service request according to the order in the queue. At the same time, if the request needs to be transmitted across domains, the access router will encapsulate the request to ensure that it can be transmitted between different networks.

[0117] 5. Response return: When the service replica instance on the computing node completes processing the request, the response data will be returned to the user end along the original path.

[0118] In some embodiments of the present invention, the steps of the method include: when the controller receives a service request for the first time, the controller parses the service type corresponding to the service request, calculates a network transmission cost matrix and an arrival rate matrix corresponding to the service type, and calculates a queue corresponding to the service type;

[0119] When the controller receives a service request again, it analyzes the service type corresponding to the service request and determines whether there is a queue for the service type currently;

[0120] If it exists, the processing task corresponding to the service request is directly sent to the corresponding cloud node or edge node based on the existing queue;

[0121] If it does not exist, construct a queue corresponding to the service type.

[0122] In some embodiments of the present invention, in the step of sending the processing task corresponding to the service request to the corresponding cloud node or edge node based on the queue, the cloud node or edge node at the head of the queue is used as the cloud node or edge node to be sent, and the cloud node or edge node is moved to the tail of the queue to update the queue.

[0123] In some embodiments of the present invention, in the step of calculating the transmission cost and propagation cost from the gateway to each cloud node and edge node based on the transmission distance and bandwidth between the gateway and the cloud node and the edge node, and calculating the network transmission cost matrix based on the transmission cost and propagation cost:

[0124] Based on the input data volume and output data volume transmitted by the gateway to each cloud node and edge node based on the preset data packet, the transmission cost is calculated based on the input data volume and output data volume and the bandwidth; the time consumed by the preset data packet from the gateway to each cloud node and edge node is used as the transmission cost;

[0125] The gateway transmits a cost value to each cloud node and edge node based on the transmission cost and the propagation cost, and constructs the cost matrix of network transmission.

[0126] In some embodiments of the present invention, in the step of calculating the transmission cost based on the input data volume and the output data volume transmitted by the gateway to each cloud node and edge node based on the preset data packet, the transmission cost is calculated based on the following formula:

[0127]

[0128] in, represents the transmission cost value from the gateway router i to the cloud node and edge node j; represents the preset input data volume of data packets for service k transmitted by gateway router i to cloud node and edge node j; represents the preset output data volume of data packets for service k transmitted by gateway router i to cloud node and edge node j; It represents the bandwidth of the preset data packet transmitted from the gateway router i to the cloud node and edge node j.

[0129] In some embodiments of the present invention, in the step of calculating the target anycast forwarding weight strategy matrix based on the arrival rate matrix and the network transmission cost matrix using the Sinkhorn algorithm, the arrival rate matrix and the network transmission cost matrix are used as two diagonal matrices of the Sinkhorn algorithm, and an objective function is constructed based on the arrival rate matrix and the network transmission cost matrix, and the final objective function value is iterated based on the Sinkhorn algorithm to obtain the target anycast forwarding weight strategy matrix.

[0130] In the specific implementation process, the Kantorovich relaxed transport problem considers how to map multiple data points in one space to multiple data points in another space, and does not believe that large-scale transport should be deterministic, but rather probabilistic transport should be considered. This probabilistic transport allows what is now commonly referred to as large-scale splitting from one source to several targets, that is, the quality at any point may be assigned to several locations. In the CPN network, the service anycast forwarding problem has changed from the traditional focus on end-to-end, one-to-one routing path problems to a planning problem that simultaneously considers service requests initiated by multiple nodes and forwards the requests to multiple different service sites. In view of the similarity in this model, the present invention uses Kantorovich relaxed transport to model the CPN service anycast forwarding problem. The basic approach of this solution is to first obtain the service instance processing capacity, and then constrain the request rate that reaches the service instance as a constraint condition of the Sinkhorn algorithm to ensure that the calculation time of the request is less than a certain threshold, and then optimize the network delay.

[0131] In some embodiments of the present invention, in the step of constructing an objective function based on the arrival rate matrix and the network transmission cost matrix, the objective function is constructed as:

[0132] L=MIN <P k , C k >

[0133] Among them, L represents the objective function value, P k represents the arrival rate matrix for service k, C krepresents the network transmission cost matrix for service k, <P k , C k > means calculating the inner product of the matrix, and MIN means minimization.

[0134] In some embodiments of the present invention, in the step of calculating the arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node, the arrival rate matrix is ​​calculated based on the following formula:

[0135] P k =α k ·X k ;

[0136] Among them, P k represents the arrival rate matrix, α k represents the arrival rate vector, X k Represents the forwarding strategy matrix.

[0137] In the specific implementation process, use n rt Indicates the number of access routers, using n cn Indicates the number of cloud nodes and edge nodes where the service is deployed, using represents the rate of user service requests per unit time for accessing CPN router i and requesting service k, then the rate vector α k It is expressed as:

[0138]

[0139] use represents the request processing capacity per unit time of service instance k of service k, then the request processing capacity vector β k It is expressed as:

[0140]

[0141] Network transmission cost matrix C k :

[0142]

[0143] definition Indicates CPN access router Forward requests for service k to the compute site The probability weight of . Then the forwarding strategy matrix X k The definition is as follows:

[0144]

[0145] For CPN access routers Remember that it will forward k to the computing site The request arrival rate is Then there is

[0146]

[0147] Kantorovich's relaxed transport model generally assumes that transport is balanced. For example, in the case of raw materials transported to the factory: the number of raw materials required by the "factory" is equal to the number of raw materials stored in the warehouse. When the two are equal, the problem model is an equality-constrained linear programming model, and it is easy to use an iterative algorithm to quickly find the model solution. When the two quantities are not equal, the transport problem becomes an unbalanced transport problem. In the unbalanced case:

[0148] 1. The total processing capacity of distributed deployed service instances is higher than the user's demand

[0149] In this case, this method assumes that there is a virtual CPN access router, and assumes that the requests collected by the router just meet the unbalanced demand. Since the router node does not really exist, and of course the additional user demand does not exist, it should not affect the overall cost band function. Setting the transmission cost to 0 can eliminate the impact of virtual requests on the optimization results.

[0150] 2. The total processing capacity of distributed deployed service instances is lower than the user's demand

[0151] When this happens, it means that the current distributed deployment solution for service k is not reasonable, and the processing capacity or the number of service replicas should be increased. Different from the warehousing and transportation problem, that is, when the inventory of raw materials is higher than the factory demand, the transportation volume of raw materials can be reduced. In CPN, when a service request arrives, the routing node must find and forward it in time, so it is impossible to forcibly reduce the request arrival rate like reducing the transportation volume of raw materials. The total processing capacity of the deployed application services does not meet the requirements. At this time, the processing capacity of the application services can be simply improved through linear simulation to match it with user needs. This method is essentially to use the processing capacity of each distributed deployed service instance as the weight to bear the user demand that exceeds the overall load capacity of the system. The coefficient γ is defined as: Adjust the processing power vector of distributed service instances The pseudo code of the service anycast strategy calculation algorithm based on optimal transmission is as follows Figure 5 shown.

[0152] In some embodiments of the present invention, in the step of calculating the network transmission cost matrix based on the transmission cost and the propagation cost, the transmission cost value transmitted by the gateway's router to each cloud node and edge node is calculated, and the propagation cost value transmitted by the gateway's router to each cloud node and edge node is measured through a preset data packet for service k. For each cloud node or edge node, the sum of the transmission cost value and the propagation cost value is calculated as the value of each position in the network transmission cost matrix to construct the network transmission cost matrix.

[0153] The beneficial effects of this program include:

[0154] 1. Computing network service anycast forwarding system architecture. The present invention proposes a service anycast forwarding architecture for computing network, including a central controller, a computing node cluster, a network topology perception module, a service quality monitoring module, and a weighted strategy forwarding queue. This architecture realizes computing resource scheduling and dynamic anycast of service requests across multiple layers of nodes by separating the control plane from the data plane, solving the problem that traditional network architectures are difficult to meet dynamic computing service requirements, and supports fine-grained balancing of differentiated service instance processing capabilities.

[0155] 2. Implementation and key function design of SDN controller under service anycast forwarding architecture. The present invention uses the RyuOpenFlow controller platform to implement the computing power network service anycast forwarding control plane function, including key functions such as service anycast forwarding entry class, real-time load status update, computing power network service anycast forwarding queue update process, and computing power network service control plane Packet_in message processing process. Through the centralized management function of the SDN controller, real-time monitoring and dynamic scheduling of cross-layer computing nodes are achieved, and a modular implementation process design is provided, so that the computing power network service anycast forwarding mechanism has the ability to be quickly deployed and flexibly expanded.

[0156] 3. Service anycast strategy model and algorithm based on Kantorovich relaxed transport. In view of the problem of fluctuation in computing resource loads of different nodes in the computing power network, load-aware forwarding based on request processing capability (RPC) is designed. In view of the problem of dynamic forwarding optimization of service requests, the present invention establishes an anycast forwarding model based on Kantorovich relaxed transport, and designs a service anycast strategy calculation algorithm based on optimal transmission, ensuring that the processing power differences and network cost differences between service instances are effectively utilized during resource scheduling, optimizing the forwarding decisions of service requests, and further improving the overall transmission efficiency and service quality of the network.

[0157] 4. Flexible centralized-distributed control architecture. The present invention realizes the organic combination of centralized control and distributed operation by separating the control plane and data plane of software-defined network (SDN). In the complex environment of cross-layer nodes, the computing network service anycast forwarding architecture makes full use of the global optimization capability of centralized control and the flexibility of the queue forwarding mechanism based on weighted strategy in the distributed data plane, enabling the system to perform fine-grained balanced scheduling of burst traffic and traffic fluctuations from the global optimal point of view, while ensuring stable service quality under high load conditions and improving resource utilization.

[0158] 5. Efficient and stable anycast strategy calculation. The present invention introduces the Kantorovich relaxed transport model and designs the optimal transmission service anycast algorithm, which balances the relationship between computing delay and network delay. The forwarding queue and request processing capacity (RPC) indicators based on the weighted strategy reduce the transmission performance fluctuation caused by load information fluctuation. In this way, even in the case of traffic fluctuations, relatively stable anycast service performance optimization can be achieved, thereby ensuring service quality.

[0159] 6. Collaborative scheduling capability across multiple layers of nodes in different domains. The computing network service anycast forwarding system architecture of the present invention has the ability to schedule resources across multiple layers and multiple domains of nodes, so that service requests can be efficiently allocated and forwarded between nodes at different levels, such as the cloud and edge. The mechanism of forwarding requests by stream is adopted, and the forwarding strategy is optimized based on queue theory to achieve collaborative work of multiple layers of nodes, which significantly reduces end-to-end latency and meets the needs of different users for low-latency and high-performance services, thereby improving the overall service quality and system stability.

[0160] An embodiment of the present invention also provides a computing power network service anycast system, which includes a computer device, wherein the computer device includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0161] The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps implemented by the aforementioned computing power network service anycast method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0162] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0163] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0164] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or replace features of other embodiments.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A computing power network service anycast method, characterized in that: This method is based on a pre-built computing network, which includes a gateway, a controller, multiple cloud nodes and multiple edge nodes. The steps of this method include: Based on the gateway receiving the service request, the gateway forwards the service request to the controller; The controller obtains the working status of each cloud node and edge node in real time, and constructs the cloud nodes and edge nodes into a node set; Based on the transmission distance and bandwidth between the gateway and the cloud nodes and edge nodes, the transmission cost and propagation cost from the gateway to each cloud node and edge node are calculated, and based on the transmission cost and propagation cost, a network transmission cost matrix is ​​calculated; Calculate the arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node; Calculate a target anycast forwarding weight strategy matrix using a Sinkhorn algorithm based on the arrival rate matrix and the network transmission cost matrix, wherein the target anycast forwarding weight strategy matrix includes a weight value of each cloud node and edge node; The number of positions of the cloud node and the edge node in the queue is determined based on the weight values ​​of the cloud node and the edge node, and a queue is constructed. Based on the queue, the processing task corresponding to the service request is sent to the corresponding cloud node or edge node.

2. The computing power network service anycast method according to claim 1, characterized in that: The method comprises the following steps: when the controller receives a service request for the first time, the controller parses the service type corresponding to the service request, calculates the network transmission cost matrix and arrival rate matrix corresponding to the service type, and calculates the queue corresponding to the service type; When the controller receives a service request again, it analyzes the service type corresponding to the service request and determines whether there is a queue for the service type currently; If it exists, the processing task corresponding to the service request is directly sent to the corresponding cloud node or edge node based on the existing queue; If it does not exist, construct a queue corresponding to the service type.

3. The computing power network service anycast method according to claim 1 or 2, characterized in that: In the step of sending the processing task corresponding to the service request to the corresponding cloud node or edge node based on the queue, the cloud node or edge node at the head of the queue is used as the cloud node or edge node to be sent, and the cloud node or edge node is moved to the tail of the queue to update the queue.

4. The computing power network service anycast method according to claim 1, characterized in that: In the step of calculating the transmission cost and propagation cost from the gateway to each cloud node and edge node based on the transmission distance and bandwidth between the gateway and the cloud node and edge node, and calculating the network transmission cost matrix based on the transmission cost and propagation cost: Based on the input data volume and output data volume transmitted by the gateway to each cloud node and edge node based on the preset data packet, the transmission cost is calculated based on the input data volume and output data volume and the bandwidth; the time consumed by the preset data packet from the gateway to each cloud node and edge node is used as the transmission cost; The gateway transmits a cost value to each cloud node and edge node based on the transmission cost and the propagation cost, and constructs the cost matrix of network transmission.

5. The computing power network service anycast method according to claim 4, characterized in that: In the step of calculating the transmission cost based on the input data volume and the output data volume and the bandwidth, based on the input data volume and the output data volume transmitted by the gateway to each cloud node and the edge node based on the preset data packet, the transmission cost is calculated based on the following formula: in, represents the transmission cost value from the gateway router i to the cloud node and edge node j; represents the preset input data volume of data packets for service k transmitted by gateway router i to cloud node and edge node j; represents the preset output data volume of data packets for service k transmitted from gateway router i to cloud node or edge node j; It represents the bandwidth of the preset data packet transmitted from the gateway router i to the cloud node or edge node j.

6. The computing power network service anycast method according to claim 1, characterized in that: In the step of calculating the target anycast forwarding weight strategy matrix using the Sinkhorn algorithm based on the arrival rate matrix and the network transmission cost matrix, the arrival rate matrix and the network transmission cost matrix are used as two diagonal matrices of the Sinkhorn algorithm, and an objective function is constructed based on the arrival rate matrix and the network transmission cost matrix. The final objective function value is iterated based on the Sinkhorn algorithm to obtain the target anycast forwarding weight strategy matrix.

7. The computing power network service anycast method according to claim 6, characterized in that: In the step of constructing an objective function based on the arrival rate matrix and the network transmission cost matrix, the objective function is constructed as: L=MIN <P k ,C k >; Among them, L represents the objective function value, P k represents the arrival rate matrix for service k, C k represents the network transmission cost matrix for service k, <P k , C k > means calculating the inner product of the matrix, and MIN means minimization.

8. The computing power network service anycast method according to claim 7, characterized in that: In the step of calculating the arrival rate matrix based on the forwarding probability weights of the gateway, the cloud node and the edge node, the arrival rate matrix is ​​calculated based on the following formula: P k =a k ·X k ; Among them, P k represents the arrival rate matrix, α k represents the arrival rate vector, X k Represents the forwarding strategy matrix.

9. The computing power network service anycast method according to any one of claims 4 to 8, characterized in that: In the step of calculating the network transmission cost matrix based on the transmission cost and the propagation cost, the transmission cost value transmitted by the gateway's router to each cloud node and edge node is calculated, and the propagation cost value transmitted by the gateway's router to each cloud node and edge node is measured through a preset data packet for service k. For each cloud node or edge node, the sum of the transmission cost value and the propagation cost value is calculated as the value of each position in the network transmission cost matrix to construct the network transmission cost matrix.

10. A computing power network service anycast system, characterized in that: The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions. The processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 9.

Images