A two-level collaborative VNF generalized orchestration method and system
By employing a two-level collaborative VNF generalization orchestration method and system, combined with deep reinforcement learning and graph neural networks, the problem of automated deployment and mapping of NFV orchestration in dynamic and heterogeneous network topologies is solved, achieving efficient VNF orchestration and quality of service assurance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-29
- Publication Date
- 2026-03-13
AI Technical Summary
Existing NFV orchestration methods struggle to achieve generalized orchestration of VNFs when faced with dynamic and heterogeneous network topology changes, resulting in performance loss and narrow applicability, and failing to meet the quality of service requirements of service level agreements.
A two-level collaborative VNF generalization orchestration method and system is adopted, which combines deep reinforcement learning and graph neural networks. Through the collaborative optimization of SFC mapping unit and VNF deployment unit, the automated deployment and mapping of VNF under diverse network topologies is realized, supporting topology generalization and service quality assurance.
It enables automated VNF orchestration in dynamic and heterogeneous network topologies, improving orchestration efficiency, responding promptly to changes in network requirements, and ensuring network service quality.
Smart Images
Figure CN115604134B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of novel network architecture technology, and in particular to a two-level collaborative VNF generalized orchestration method and system. Background Technology
[0002] Network Function Virtualization (NFV) transforms traditional networks by decoupling network functions from hardware using virtualization and cloud technologies, and abstracting network services into software called Virtual Network Functions (VNFs) running on basic hardware. In the context of NFV and agile service production, resource virtualization and Service Function Chains (SFCs) play a crucial role in automating network service deployment. NFV orchestration frameworks address the challenges of designing NFV management and orchestration to provide dynamic SFC-as-a-service. However, current NFV orchestration relies on varying constraints for different SFC requests (SFCRs), network topology, traffic load, etc. When parameter settings are mismatched, NFV orchestration may suffer performance penalties.
[0003] In recent years, NFV orchestration has attracted industry attention. Ensuring the agility, automation, and generalization capabilities of NFV orchestration has become a hot research topic in the field of new network architectures. However, current research largely focuses on two stages of VNF orchestration: first, deploying VNFs, and then mapping incoming SFC requests to the physical network based on the deployed VNF distribution map. This approach only considers VNF orchestration in fixed network topologies, neglecting dynamic changes in the current network topology or VNF orchestration in different network topologies. It struggles to handle heterogeneous networks and effectively achieve generalized VNF orchestration, thus limiting its applicability. Summary of the Invention
[0004] To address the complexities of parameter-environment matching in NFV orchestration, this invention designs a two-level collaborative VNF generalized orchestration method and system. This system adapts to the large number of uncertain and volatile SFC requests and diverse network topologies in NFV networks, automating the NFV orchestration process while ensuring compliance with Service Level Agreement (SLA) quality of service levels. This invention incorporates the concept of Network Function Virtualization (NFV), supporting automated VNF deployment and SFC mapping in NFV networks. It also supports collaborative optimization of both the VNF deployment and SFC mapping phases and generalized orchestration across different network topologies.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] This invention discloses a two-level collaborative VNF generalization orchestration system, including an SFC mapping unit and a VNF deployment unit; the SFC mapping unit includes a mapping controller; the VNF deployment unit includes a deployment controller; both the mapping controller and the deployment controller include a topology generalization module.
[0007] The SFC mapping unit is used to find the shortest path that can handle the current SFC request in the physical network whose topology may change, based on the current VNF distribution, and connect the corresponding physical network nodes in sequence according to the direction of the path.
[0008] The VNF deployment unit is used to decide the distribution of VNF instances in the physical network of the current topology based on the current network status and SFC mapping quality.
[0009] The mapping controller is used to manage the SFC mapping unit for physical node selection;
[0010] The deployment controller is used to control the VNF deployment unit, replan the distribution of VNFs on each physical node, and make decisions on VNF deployment based on network status and SFC sequence as input and SFC mapping quality as feedback.
[0011] The topology generalization module is used to solve the VNF orchestration problem using graph reinforcement learning.
[0012] Furthermore, both the mapping controller and the deployment controller are based on intelligent algorithms that can achieve topology generalization and have the ability to adapt to different network topologies for VNF orchestration.
[0013] Furthermore, the SFC mapping unit and the VNF deployment unit interact with each other at the granularity of an SFC request.
[0014] Furthermore, the mapping quality of an SFC is quantified as the ratio of the latency of its mapping logical link to the latency of the shortest path between its source and destination nodes.
[0015] Furthermore, the topology generalization module employs a combination of deep reinforcement learning and graph neural networks to achieve topology generalization.
[0016] Furthermore, in the topology generalization module, one physical link corresponds to two directed edges in the graph neural network.
[0017] Furthermore, in the topology generalization module, within the graph neural network:
[0018] The input feature value of the node is input_node_feature.
[0019] input_node_feature=("resource","VNF","ID")
[0020] Wherein, resource represents the current remaining amount of resources on the node, VNF represents the type of VNF deployment on the node, and ID represents the node number;
[0021] The input feature value of the edge is input_edge_feature.
[0022] input_edge_feature=("delay","bandwidth")
[0023] Where delay and bandwidth represent the transmission delay and bandwidth on the link, respectively;
[0024] The global input feature value is input_global_feature.
[0025] input_global_feature=("start","end","duration","SFC")
[0026] Where duration, start, and end represent the working duration, source node number, and destination node number of the SFC to be mapped, respectively; SFC represents other information in the SFC request mapping interaction granularity fi, i.e. Encoding into a tensor T of dimension (n, n, 2), where the two values of the third dimension of T represent... Capacity requirements for nodes and links Slicing T along its third dimension yields an n×n permutation matrix A. A transforms the fixed-order VNF tuples into VNF tuples in fi, and then merges them. Together with A, we obtain the tensor T.
[0027] Another aspect of this invention proposes a two-level cooperative VNF generalization orchestration method, comprising:
[0028] The deployment of VNFs and mapping of SFCs are performed alternately by the SFC mapping unit and the VNF deployment unit, thereby automatically adjusting the orchestration scheme of VNFs.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] This invention enables automated VNF orchestration with topology generalization, effectively improving VNF orchestration efficiency, facilitating network administrators to respond promptly to changes in network requirements, and ensuring network service quality. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the architecture of a two-level collaborative VNF generalized orchestration system according to an embodiment of the present invention;
[0032] Figure 2 A functional diagram illustrating a two-level collaborative VNF generalization orchestration system provided in an embodiment of the present invention;
[0033] Figure 3 This is a schematic diagram of the SFC mapping principle provided in an embodiment of the present invention;
[0034] Figure 4 This is an example diagram of an SFC request on a certain topology provided in an embodiment of the present invention;
[0035] Figure 5 This is a schematic diagram of VNF deployment provided in an embodiment of the present invention. Detailed Implementation
[0036] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:
[0037] This invention proposes a two-level cooperative VNF generalized orchestration system, such as... Figure 1 As shown. From a macro perspective, the overall design of this system is based on the idea of decoupling hardware network devices from network functions through network function virtualization. It achieves automated network function orchestration and global optimization that adapts to diverse network topologies through intelligent algorithms for VNF deployment and SFC mapping.
[0038] Furthermore, the upper layer is the VNF deployment unit, and the lower layer is the SFC mapping unit. The VNF deployment unit is responsible for placing different types of VNF instances based on predicting user service needs, while the SFC mapping unit is responsible for mapping these VNF instances to the physical network and connecting them in a certain order to meet the user's network function requirements. The VNF deployment unit can influence the mapping quality of the SFC mapping unit, and the mapping results of the SFC mapping unit, as feedback, also affect the VNF deployment unit. Therefore, the two need to cooperate to achieve better orchestration. The SFC mapping unit is managed by the mapping controller, which interacts at the granularity of an SFC request, processing incoming requests sequentially until mapping can no longer be successful.
[0039] Furthermore, both the mapping controller and the deployment controller are implemented based on intelligent algorithms that can achieve topology generalization, and both have the ability to adapt to different network topologies for VNF orchestration. That is, both the mapping controller and the deployment controller include a topology generalization module.
[0040] This system can use artificial intelligence algorithms such as graph reinforcement learning to solve the VNF orchestration problem for the following reasons: 1) VNF orchestration can be expressed as a Markov process, which is suitable for optimization by deep reinforcement learning (DRL); 2) Graph neural networks (GNNs), as deep neural networks of DRL, can receive and output data of variable dimensions, which is suitable for network scenarios with diverse topologies.
[0041] For example, the granularity of the SFC request mapping interaction mentioned above can be f1, f2, ..., f T1 This is represented by a cycle called a DRL. The VNF deployment unit is managed by the deployment controller, which interacts at a granularity of one cycle, for example, cycle 0, cycle 1, ..., which is an infinite sequence of DRLs.
[0042] The key to topology generalization is using a Generative Neural Network (GNN) as the "depth" in a Graph Deep Reinforcement Learning (DRL), known as Generalized (or Graph) Deep Reinforcement Learning (GDRL). GNNs operate on graphs and construct their computations accordingly, which can be mapped to the network's topology. When the topology of the target network to be optimized changes, only the topology of the GNN in the GDRL needs to be adjusted accordingly, without changing the dimensions of the neural network input data (padding or pruning), to continue applying the system for VNF orchestration. Based on this, the inputs and outputs of the VNF orchestration problem are reasonably mapped to (s, a, r) in the DRL, and then encoded as features of the GNN. Throughout the process, the user only needs to obtain the virtual network view generated by the system, without needing to know the underlying network information, and changes to the underlying network will not affect the user experience.
[0043] To facilitate a further understanding of the present invention, the following will be combined with... Figure 2 The specific embodiments of the present invention will be described in detail below:
[0044] The SFC mapping unit, as described in this invention, loads an SFC mapping unit between the physical network and the VNF deployment unit. Figure 3 For example, the number of VNF types that need to be deployed is usually constant, and their total number is denoted as n. The SFC mapping unit is used to determine the physical node on which each VNF in the service request should run. When the remaining computing resources on some physical nodes meet the needs of the VNFs, and the remaining bandwidth resources on the physical links meet the data transmission needs between the VNFs, the mapping controller running in the SFC mapping unit comprehensively considers the quality of service (QoS) that these candidate nodes can obtain and selects the nodes accordingly.
[0045] Quality of Service (QoS) is a relative quantity, specifically the ratio of the QoS of an actual orchestrated SFC to its theoretically optimal value. There are multiple metrics for QoS; here we focus only on link latency, which is most relevant to VNF orchestration. Therefore, the mapping quality of an SFC is quantified as the ratio of the latency of its mapped logical link to the latency of the shortest path between its source and destination nodes. The VNF orchestration quality for one cycle is defined as the average mapping quality of all SFCs, including active and inactive SFCs.
[0046] The present invention adds a VNF deployment unit on top of the SFC mapping unit. During the SFC mapping process described in the SFC mapping unit, we assume that VNF deployment is fixed until the end of a cycle that enables topology-generalized SFC mapping. Before SFC mapping begins, the deployment controller can re-plan the distribution of VNFs on each physical node to facilitate better SFC mapping in the next cycle. Similarly, the deployment controller uses network status and SFC sequence as input, and SFC mapping quality as feedback, to decide on VNF deployment. It is important to note that since future SFC requests are unknown, the deployment controller can only use historical SFC requests as input (an example of SFC requests on a certain topology is shown below). Figure 4 As shown in the diagram, it is necessary to forecast future demand. A VNF deployment illustration is shown below. Figure 5 As shown.
[0047] The topology generalization module is loaded into both the SFC mapping unit and the VNF deployment unit in this invention to enable wide application in diverse network topology scenarios. The topology of the neural network (GNN) in GDRL is the same as the topology of the physical network, where one physical link corresponds to two directed edges in the GNN. Therefore, changes in network topology do not affect the operation of GNN / GDRL, which is a crucial property for optimization algorithms in dynamic networks. However, the nodes, edges, and global neural networks in GNN use fully connected neural networks, which can only process data of fixed dimensions. We need to encode the data in the network as the rule input for GDRL and decode its output as diverse mapping strategies, while ensuring that the data organization has the logical relationships that GNN can handle.
[0048] In GNNs, node, edge, and global input feature values represent information containing the network state and the SFC to be mapped. The input feature values at nodes are...
[0049] input_node_feature=("resource","VNF","ID")
[0050] Here, `resource` represents the current remaining resources on the node, encoded into a tensor of shape (1), i.e., a scalar of dimension (1). `VNF` represents the type of VNF deployment on the node, encoded into a tensor of shape (n), i.e., a scalar of dimension (n). For example, if the current network's VNF tuple order is (VNF0, VNF1, VNF2, VNF3, VNF4), and network node 0 has VNF0 and VNF2 deployed, then its VNF is (1,0,1,0,0). `ID` represents the node's ID, which is also a scalar.
[0051] The input feature value of the edge in GNN is
[0052] input_edge_feature=("delay","bandwidth")
[0053] delay and bandwidth represent the transmission delay and bandwidth on the link, respectively, both of which are scalars.
[0054] The global input feature value in GNN is
[0055] input_global_feature=("start","end","duration","SFC")
[0056] Here, duration, start, and end represent the working duration, source node number, and destination node number of the SFC to be mapped, respectively, all of which are scalars. The SFC needs to fully represent the other information in fi, that is... Encode into a tensor T of shape (n, n, 2). The two values of the third dimension of T represent... Capacity requirements for nodes and links and Slicing T along its third dimension yields an n×n matrix, denoted as A. A is a permutation matrix that transforms a fixed-ordered set of VNF tuples into VNF tuples in fi. For example,
[0057]
[0058] Therefore, merger The tensor T can be obtained by combining A and E. Taking equation (1) as an example, then...
[0059]
[0060] Based on the above embodiments, the present invention also discloses a two-level cooperative VNF generalization orchestration method, comprising:
[0061] The deployment of VNFs and mapping of SFCs are performed alternately by the SFC mapping unit and the VNF deployment unit, thereby automatically adjusting the orchestration scheme of VNFs.
[0062] In summary, through the cooperation of the above modules, this invention can achieve automated VNF orchestration with topology generalization, effectively improve VNF orchestration efficiency, facilitate network administrators to respond to changes in network requirements in a timely manner, and ensure network service quality.
[0063] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A two-stage cooperative VNF generalization orchestration system, characterized in that, The SFC mapping unit comprises a mapping controller, and the VNF deployment unit comprises a deployment controller; the mapping controller and the deployment controller both comprise a topology generalization module; The SFC mapping unit is configured to find a shortest path capable of processing a current SFC request in a physical network in which a topology structure can change according to a current VNF distribution, and sequentially connect corresponding physical network nodes in the order of the path; The VNF deployment unit is configured to decide the distribution of VNF instances in the current topology structure of the physical network according to a current network state and SFC mapping quality; The mapping controller is configured to manage the SFC mapping unit to select physical nodes; The deployment controller is configured to control the VNF deployment unit to re-plan the distribution of VNFs on each physical node, and to use a network state and an SFC sequence as input and use SFC mapping quality as feedback to decide the deployment of VNFs; The topology generalization module is configured to use graph reinforcement learning to solve the VNF orchestration problem. In the graph neural network in the topology generalization module, a global input feature is used to encode SFC request information, and the global input feature is constructed as a tensor T with a dimension of (n, n, 2), where n is the total number of VNF types; two values in the third dimension of the tensor T represent the capacity requirements of VNFs for nodes and links, respectively; slicing the tensor T along the third dimension obtains a permutation matrix A with a size of n x n, and the permutation matrix A is used to convert a VNF tuple in a fixed order into a VNF tuple sequence in a current SFC request. 2.The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, Both the mapping controller and the deployment controller are based on an intelligent algorithm that can realize topology generalization, and both have the ability to adapt to different network topologies for VNF orchestration. 3.The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, Both the SFC mapping unit and the VNF deployment unit interact with each other in the granularity of a SFC request. 4.The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, The mapping quality of a SFC is quantified as the ratio of the delay of the mapping logical link to the delay of the shortest path between the source node and the destination node. 5.The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, In the topology generalization module, topology generalization is realized by combining deep reinforcement learning with a graph neural network.
6. The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, In the topology generalization module, one physical link corresponds to two directed edges in the graph neural network.
7. The two-stage cooperative VNF generalization orchestration system of claim 1, wherein, In the topology generalization module, in the graph neural network: input feature values on the nodes for where resource represents the current resource remaining amount on the node, VNF represents the VNF deployment type on the node, and ID represents the number on the node. input feature values on the edges for wherein, , respectively represent the transmission delay and bandwidth on the link. Global input feature values To Wherein, duration, start and end respectively represent the working duration of the SFC to be mapped, the source node number, and the destination node number; SFC represents other information in the SFC request mapping interaction granularity fi, that is , encoded into a tensor T with dimensions (n, n, 2), the two values of the third dimension in T respectively represent The capacity requirements of nodes and links 、 T can be sliced with the third dimension to obtain a permutation matrix A with a size of n x n, A converts the VNF tuple in the fixed order VNF tuple fi, and combines and A to obtain the tensor T. 8.A two-level collaborative VNF generic orchestration method based on the two-level collaborative VNF generic orchestration system according to any one of claims 1-7, characterized in that, The method comprises the following steps: The SFC mapping unit and the VNF deployment unit are used to alternately deploy VNFs and map SFCs, and to automatically adjust the VNF orchestration scheme. In the decision process of the VNF deployment unit, the graph neural network in the topology generalization module receives SFC request information encoded as a specific tensor T and a permutation matrix A as global input.