Enterprise-level Ultra-large-scale networking methods and systems based on SDN and NFV

Abstract

This invention relates to the field of cloud computing platform networking technology, and particularly to enterprise-level ultra-large-scale networking methods and systems based on SDN and NFV. The invention proposes the following scheme: constructing a joint resource scheduling plane, using two-dimensional coordinate mapping to map resources and demands, calculating resource matching degree, determining the global resource domain, and matching the physical computing, storage, and network resources of VNF instances and physical nodes based on a hierarchical decomposition VNF deployment optimization model. Furthermore, a dynamic computing power migration algorithm is employed, optimizing load balancing through an A* search strategy, and combining this with a distributed probe cluster to monitor the VNF instance running status in real time, dynamically adjusting resource allocation. To address resource conflict issues, a demand rectangle splitting mechanism is provided, dividing computing tasks into sub-tasks and optimizing deployment to improve computing throughput.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing platform networking, and particularly relates to an enterprise-level super-large-scale networking method and system based on SDN and NFV. BACKGROUND

[0002] The existing enterprise-level super-large-scale networking method has many challenges in VNF instance deployment optimization, computing resource scheduling and load balancing. The traditional method mainly relies on static resource allocation and SLA-based rule matching, which is difficult to cope with dynamic changes in resource demand, high concurrency of computing tasks and resource adaptation in heterogeneous environments, resulting in low utilization of computing resources, high task migration cost and large network transmission overhead. In addition, the existing VNF deployment strategy is difficult to efficiently handle the coordination and optimization of computing, storage and network resources in a super-large-scale cloud computing platform. When computing resources are scarce or task concurrency increases, resource contention, computing node overload and unbalanced task scheduling may occur. Especially in AI computing, cloud storage and high-performance computing (HPC) scenarios, the computing demands of different VNF instances differ greatly, and traditional scheduling strategies cannot accurately match, resulting in serious resource fragmentation and affecting system throughput.

[0003] The above existing technologies all have the problem proposed in the background: different businesses have different demands for computing, storage and network resources, resulting in VNF placement optimization being an NP-hard problem. To solve the above problems, the present application designs an enterprise-level super-large-scale networking method and system based on SDN and NFV. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the deficiencies of the prior art and provide an enterprise-level super-large-scale networking method and system based on SDN and NFV. A resource joint scheduling plane is constructed, and two-dimensional coordinate mapping is used to map resources and demands, calculate resource matching degree, and determine the global resource domain. Based on a hierarchical decomposition VNF deployment optimization model, the physical computing, storage and network resources of VNF instances and physical nodes are matched. Further, a dynamic computing power migration algorithm is used to optimize load balancing through an A* search strategy, and a distributed probe cluster is used to monitor the running state of VNF instances in real time, and dynamically adjust resource allocation. To solve the resource conflict problem, a demand rectangle splitting mechanism is provided to split computing tasks into subtasks and optimize deployment, thereby improving computing throughput.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] The enterprise-level super-large networking method based on SDN and NFV is applied to a network service component of a cloud computing platform. The super-large networking needs to satisfy that the cloud computing platform manages more than or equal to 10000 virtualized network devices and concurrently runs more than or equal to 1000 heterogeneous VNF instances. The enterprise-level super-large networking method comprises the following steps:

[0007] A resource joint scheduling plane is constructed by the SDN and the NFV.

[0008] According to the resource joint scheduling plane, the heterogeneous VNF instances are matched to generate a deployment decision of the heterogeneous VNF instances.

[0009] A distributed probe cluster is deployed in the network service component, and the deployment decision is updated in real time to perform super-large networking for an enterprise. The probe cluster is used to monitor running data of the VNF instances.

[0010] The resource joint scheduling plane is constructed by the SDN and the NFV, and comprises the following steps:

[0011] An SDN controller collects physical resource indexes of the virtualized network devices based on an extended OpenFlow protocol, wherein the physical resource indexes comprise physical computing resources, storage resources and network resources, and stores the physical resource indexes in a resource state database.

[0012] An NFV orchestrator analyzes a requirement description file submitted by an enterprise user and analyzes requirement characteristics, wherein the requirement characteristics comprise computing resource requirements, storage requirements and network performance requirements.

[0013] The physical resource indexes and the requirement characteristics are connected through a southbound interface and a KVM virtualization layer of a cloud platform to generate the resource joint scheduling plane.

[0014] The physical resource indexes and the requirement characteristics are connected, and the connection comprises the following steps:

[0015] The physical resource indexes and the requirement characteristics are associated and mapped to obtain a resource mapping relationship, wherein the physical computing resources are associated with the computing resource requirements, the storage resources are associated with the storage requirements, and the network resources are associated with the network performance requirements.

[0016] According to the resource mapping relationship, a resource adaptation matrix is constructed through a predefined resource adaptation rule. The resource adaptation matrix standardizes resource states of each physical node to form a two-dimensional adaptation matrix of the physical resource indexes and the requirement characteristics. The resource adaptation matrix is used to describe adaptation degrees of different physical nodes in each requirement dimension.

[0017] matching with the heterogeneous VNF instances, generating a deployment decision of the heterogeneous VNF instances, comprising:

[0018] According to the resource joint scheduling plane, a VNF deployment optimization model based on hierarchical decomposition is constructed, and a matching result is generated, wherein the VNF deployment optimization model includes global resource domain planning and local node deployment optimization, the global resource domain planning is used to determine the target resource domain of the VNF instance, and the local node deployment optimization is used to match the VNF instance with the physical computing, storage and network resources of the physical node according to the target resource domain;

[0019] The matching result is updated by a dynamic computing power migration algorithm to generate a deployment decision.

[0020] The global resource domain planning includes:

[0021] According to the resource joint scheduling plane, a resource and demand plane coordinate graph is established in a two-dimensional coordinate system, wherein the horizontal axis represents the physical resource computing power distribution, and the vertical axis represents the demand characteristics;

[0022] The physical node is mapped to a resource polygon, the demand of the VNF instance is taken as a demand node, and the demand node is mapped to a demand rectangle;

[0023] The overlap degree of the resource polygon and the demand rectangle is calculated, and the resource polygon with an overlap degree greater than an overlap threshold is taken as the target resource domain of the corresponding demand node;

[0024] Each heterogeneous VNF is split into multiple functional components, and synchronous communication frequency is defined between the components, wherein each functional component corresponds to different demand characteristics;

[0025] In each target resource domain, the functional components are allocated to the demand nodes with the highest similarity in demand characteristics, wherein the similarity is calculated by cosine similarity;

[0026] If a single node cannot meet all resource demands of the components, the functional components are split into sub-components that can run across nodes in cooperation, and a refined network flow table is issued by the SDN controller.

[0027] The local node deployment optimization includes:

[0028] On each demand node, a running container is constructed, wherein the running container includes an initial container and a standby container, the initial container is used to carry the functional components currently being executed, and the standby container is in an idle state;

[0029] According to the running condition of the running container, it is judged whether the demand node has a secondary overlap state, wherein the secondary overlap state refers to the competition conflict between demand nodes;

[0030] When the secondary overlap state exists, the secondary overlap part is reassigned.

[0031] The reassignment of the secondary overlap part includes:

[0032] A conflict factor is calculated according to a change trend of a corresponding demand rectangle in a resource joint scheduling plane;

[0033] If the conflict factor is less than a preset conflict threshold, a running container maintains a current running state;

[0034] If the conflict factor is greater than or equal to the preset conflict threshold, the standby container is allocated to a demand node where the secondary overlap state exists;

[0035] When the standby container is insufficient to meet a demand of the demand node where the secondary overlap state exists, a demand rectangle splitting mechanism is used to split a function component currently in conflict into multiple independent subtasks on each demand node, and the function component is reassigned according to a priority.

[0036] A dynamic computing power migration algorithm is used to update a matching result, including:

[0037] A load balancing degree of a current resource polygon is calculated according to the matching result;

[0038] If the load balancing degree is greater than a preset load threshold, a computing power migration demand is triggered;

[0039] A computing power migration factor of a target resource domain where the computing power migration demand exists is calculated, and the computing power migration factor includes a computing power overflow degree, a data transmission cost, and a migration time estimation;

[0040] The computing power migration factor is matched by using an A* algorithm according to the resource joint scheduling plane, to obtain a second target domain, and the A* algorithm calculates a migration path according to a topology of a remaining idle resource domain of the resource joint scheduling plane, and selects an idle resource domain with a minimum calculation cost;

[0041] Migration is performed to the second target domain, and a resource joint scheduling plane is updated according to a load status of a target node after the computing power migration, a matching relationship between a resource polygon and a demand rectangle is recalculated, and a deployment decision is generated.

[0042] An enterprise-level super-large-scale networking system based on SDN and NFV, the system including a data acquisition module, a plane construction module, and an instance deployment module;

[0043] The data acquisition module is configured to acquire physical resource indicators and a demand description file;

[0044] The plane construction module is configured to process the physical resource index and the demand description file through the SDN and the NFV, and construct a resource joint scheduling plane.

[0045] The instance deployment module is configured to match the VNF instance according to the resource joint scheduling plane, and generate a deployment decision of the VNF instance.

[0046] The instance deployment module comprises:

[0047] A global planning unit is configured to determine a target resource domain of the VNF instance.

[0048] A local optimization unit is configured to match the VNF instance with the physical computing, storage and network resources of the physical node according to the target resource domain.

[0049] A matching updating unit is configured to update the matching result through a dynamic computing power migration algorithm, and generate the deployment decision.

[0050] Compared with the prior art, the present application has the following beneficial effects:

[0051] The present application effectively improves the resource scheduling efficiency and the computing task throughput capacity in the super large scale cloud computing environment through the resource joint scheduling plane, the hierarchical decomposition VNF deployment optimization model, the dynamic computing power migration algorithm and the demand rectangular splitting mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0052] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-restrictive embodiments, made with reference to the accompanying drawings:

[0053] Figure 1 Fig. 1 is a flowchart of an enterprise-level super large scale networking method based on SDN and NFV according to Embodiment 1 of the present application;

[0054] Figure 2 Fig. 2 is a schematic diagram of target resource domain selection according to Embodiment 1 of the present application;

[0055] Figure 3 Fig. 3 is a schematic diagram of an enterprise-level super large scale networking system based on SDN and NFV according to Embodiment 2 of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application.

[0057] Embodiment 1

[0058] Please refer to Figure 1In one embodiment of the present application, an enterprise-level super-large networking method based on SDN and NFV is applied to a network service component of a cloud computing platform. The super-large networking method needs to satisfy the management of more than or equal to 10,000 virtualized network devices and the concurrent operation of more than or equal to 1,000 heterogeneous VNF instances of the cloud computing platform. The present application optimizes the matching and deployment of VNF instances through the construction of a resource joint scheduling plane, and realizes the optimized management of enterprise-level super-large networking in combination with dynamic computing power migration and a distributed probe cluster. The specific steps of the enterprise-level super-large networking method are as follows:

[0059] S1: Constructing a resource joint scheduling plane through SDN and NFV;

[0060] In this embodiment, the construction of the resource joint scheduling plane adopts a two-dimensional mapping method of physical resource indicators and demand characteristics. Real-time data of physical computing resources, storage resources and network resources are collected by an SDN controller, and the demand characteristics of a service are analyzed by an NFV orchestrator.

[0061] This construction method can intuitively display the resource supply and demand, convert the deployment optimization problem of VNF instances into a geometric matching problem, reduce the computational complexity of the NP-hard problem, improve the global resource utilization efficiency, and reduce the fragmentation of computing resources.

[0062] S2: Matching the heterogeneous VNF instances according to the resource joint scheduling plane to generate a deployment decision of the heterogeneous VNF instances;

[0063] In this embodiment, a VNF deployment optimization model based on hierarchical decomposition is constructed to first plan a global resource domain and determine the target resource domain of the VNF instances, and then optimize the local node deployment to select appropriate computing nodes for specific deployment. The VNF instances are first split into multiple functional components, each component corresponding to different computing resource requirements, and synchronous communication frequencies are defined between the components. In the target resource domain, the functional components are matched to the physical nodes that best meet their demand characteristics by using cosine similarity calculation. If a single node cannot meet all resource requirements, the components are split into multiple sub-tasks through a cross-node cooperative operation strategy, and detailed flow table rules are issued by the SDN controller to ensure the optimal data path. In addition, after the matching is completed, the deployment decision is adjusted according to a dynamic computing power migration algorithm, the A* algorithm is used to calculate the optimal computing power migration path, and the VNF instances with high load are migrated to target resource domains with low computing power load.

[0064] Through two-dimensional visualization of resource matching, precise mapping of functional components and dynamic computing power migration mechanism, the global optimization capability of VNF instance deployment is improved, the efficient utilization of heterogeneous computing resources is ensured, and the problem of uneven load of computing nodes is reduced.

[0065] S3: deploying a distributed probe cluster in a network service component, updating a deployment decision in real time to network a super-large scale enterprise;

[0066] In this embodiment, the distributed probe cluster is used to monitor the running state of the VNF instance and update the resource joint scheduling plane in real time. The probe cluster is deployed in different computing nodes, and each probe is responsible for monitoring the computing load, storage I / O throughput and network transmission delay of multiple VNF instances, and periodically uploading the collected data to the resource state database. The resource scheduling is changed from static planning to real-time adaptive optimization, which greatly improves the accuracy of computing power scheduling and the resource utilization of computing nodes, ensures the stable operation of super-large scale VNF instances, and reduces the waste of computing resources.

[0067] In the prior art, the deployment optimization of enterprise-level super-large scale networking VNF instances is a typical NP-hard problem. Due to the dynamic changes of computing resources, storage resources and network resources and the heterogeneity of VNF instances, traditional deployment optimization methods are difficult to provide efficient and accurate resource allocation schemes in a large-scale environment. Existing VNF instance deployment methods mainly rely on heuristic scheduling or rule matching mechanisms. These methods can provide certain optimization effect in small-scale scenarios, but in the case of resource heterogeneity, dynamic changes of business demand and high concurrency of VNF instance running, they lack intuitive display of resource distribution, making it difficult for scheduling strategies to adapt to rapid changes in business load, thereby affecting overall computing efficiency and network performance.

[0068] In this embodiment, the present application proposes an enterprise-level super-large scale networking method based on a resource joint scheduling plane. Taking a cloud computing platform as an example, the method uses two-dimensional resource matching modeling, dynamic computing power migration and adaptive local optimization to make the deployment process of VNF instances more efficient and accurate, and improve resource utilization.

[0069] Specifically, the resource joint scheduling plane is used to map the states of computing resources, storage resources and network resources in two dimensions. By representing the physical resource computing power distribution and demand characteristics in a two-dimensional coordinate system, intuitive visual management of resource supply and demand is achieved. For example, in a super-large scale data center, some computing nodes have high load, and traditional methods are difficult to quickly determine which computing resources can be migrated. However, through the method of the present application, all computing resources are represented in the form of resource polygons on the resource joint scheduling plane, and the demand of VNF instances is represented in the form of demand rectangles. The computing overlap degree can intuitively reflect the resource adaptation. This two-dimensional visual management method converts the resource scheduling problem into a geometric matching problem, reduces the computational complexity, and enables the scheduling system to complete the resource matching calculation of tens of thousands of VNF instances within milliseconds, effectively solving the resource matching problem in super-large scale networking.

[0070] Preferably, the enterprise-level super-large networking method based on SDN and NFV proposed in this embodiment can also be applied to a computing power scheduling platform, and is particularly suitable for scheduling optimization of heterogeneous GPU computing resources, so as to solve the computing power matching and load balancing problems of different types of computing devices in a super-large computing environment.

[0071] Specifically, in the computing power scheduling platform, the computing resources are usually composed of GPUs of multiple architectures (such as NVIDIA A100, RTX 4090, and domestic GPUs). Different types of GPUs have different computing capabilities, power consumption characteristics, and suitable task types. Traditional computing power scheduling methods are usually based on static resource allocation strategies, which result in low computing resource utilization, long scheduling response time, and waste of GPU resources due to task mismatching. In this embodiment, by constructing a resource joint scheduling plane, the GPU computing resources can be mapped to a resource polygon, and the computing task demand can be mapped to a demand rectangle. The optimal computing power allocation scheme is automatically determined by calculating the overlap degree, so as to ensure the accurate matching of task scheduling and GPU computing capability. In addition, for high-load tasks or AI training tasks, a dynamic computing power migration algorithm is adopted. The optimal migration path is calculated by an A* search strategy, and the computing task is automatically migrated from a low-computing-power node to a high-performance GPU node, so as to optimize the throughput and execution efficiency of the computing task. For the computing resource contention problem in a heterogeneous GPU environment, a distributed probe cluster can be used to monitor the computing power occupation in real time, and a secondary overlap state detection can be triggered to dynamically adjust the GPU task mapping relationship, so as to prevent the occurrence of computing power bottlenecks, thereby providing a more efficient and intelligent computing power scheduling scheme for large-scale AI training, cloud computing, and high-performance computing (HPC).

[0072] The specific steps of S1 are as follows:

[0073] S1.1: The SDN controller collects physical resource indicators based on the extended OpenFlow protocol, wherein the physical resource indicators include physical computing resources, storage resources, and network resources, and stores the physical resource indicators in a resource state database;

[0074] In this embodiment, the SDN controller extends the traditional OpenFlow protocol to support multi-dimensional resource state monitoring in a super-large cloud computing platform. The basic function of the OpenFlow protocol is to manage and control the forwarding of data flow, but its native function mainly focuses on the flow scheduling of the network layer. In this application, the real-time monitoring capability of computing resources and storage resources is added based on OpenFlow to meet the dynamic scheduling requirements of resources in a cloud computing environment.

[0075] Specifically, the SDN controller deploys a resource monitoring agent (RMA) on each computing node, which collects CPU occupancy, GPU utilization, storage I / O throughput, and memory usage based on periodic polling and event triggering mechanisms, and encapsulates the data into a standard flow table through an extended OpenFlow protocol. To improve the efficiency of resource data collection, the SDN controller uses traffic telemetry technology based on NetFlow, sFlow, and IPFIX protocols, and deploys hardware traffic probes on switch and server ports to accurately obtain bandwidth occupancy, packet loss rate, and latency information of each computing node.

[0076] Further, to avoid the impact of data collection on business traffic, the SDN controller uses a hierarchical sampling strategy, i.e., for low-load network environments, it uses a large time interval for collection, while for high-load networks, it shortens the sampling interval to provide more accurate resource usage status. All collected physical resource indicators are stored in a resource state database (RSD), which uses a time-series data storage architecture to support high-throughput resource data writing and allow the orchestrator to predict resource utilization based on historical trends. By comprehensively sensing the physical computing, storage, and network resource status, the subsequent resource joint scheduling plane can be accurately optimized based on the real resource distribution, thereby improving the rationality of VNF instance deployment, reducing resource waste on computing nodes, and improving computing task throughput efficiency.

[0077] S1.2: The NFV orchestrator parses the requirement description file submitted by the enterprise user and analyzes the requirement characteristics, wherein the requirement characteristics include computing resource requirements, storage requirements, and network performance requirements;

[0078] In this embodiment, the core task of the NFV orchestrator is to parse the requirement description file (SDF) submitted by the enterprise user, which is in JSON / YAML format and contains information such as user computing resource requirements, storage resource requirements, and network performance requirements. To ensure compatibility of different business scenarios, the NFV orchestrator supports automatic conversion of VNFD (VNF Descriptor) of ETSI NFV-MANO standard, and extracts the computing power requirements, storage requirements, and bandwidth requirements from it for resource adaptation.

[0079] Specifically, the orchestrator adopts a multi-level parsing engine, first extracts the computing specifications (CPU core number, GPU acceleration type, memory capacity), storage requirements (local SSD, remote block storage or object storage), and network requirements (minimum bandwidth, maximum latency, QoS guarantee level) in the SDF file based on regular matching and context analysis, and then enters the semantic matching stage. Based on the resource adaptation model constructed based on historical deployment data, it is determined whether additional optimization is required. For example, when the computing requirements of a VNF instance include "supporting low-latency AI inference tasks", the orchestrator will automatically filter GPU clusters with NVLink interconnection in the resource joint scheduling plane to ensure that the computing nodes can meet the performance requirements of low-latency inference.

[0080] Further, the NFV orchestrator also integrates a hierarchical demand matching algorithm, which constructs a KNN (K-Nearest Neighbor) similarity model for computing, storage, and network requirements, and finds the closest existing resource configuration in the resource library to reduce unnecessary computing power migration and reduce computing resource fragmentation. This approach ensures that computing resources, storage resources, and network resources are accurately matched to the requirements of VNF instances, improving overall resource utilization and reducing performance degradation caused by resource bottlenecks.

[0081] S1.3: Through the Southbound interface with the KVM virtualization layer of the cloud platform, the physical resource indicators are connected with the demand characteristics to generate a resource joint scheduling plane;

[0082] In this embodiment, the NFV orchestrator needs to interact with the KVM virtualization layer of the cloud computing platform through the Southbound interface (Southbound API) to complete the mapping of resources and demands and generate a resource joint scheduling plane. KVM (Kernel-based Virtual Machine) is a virtualization module of the Linux kernel that provides efficient virtual machine management functions. In this application, the NFV orchestrator interacts with KVM through Libvirt API to obtain virtual machine instance information, NUMA node topology, CPU binding relationship, and I / O channel status on the physical server. Since KVM itself does not have unified resource scheduling capabilities, this embodiment introduces a resource isolation technology based on cgroup (control group), allowing computing resources to be dynamically adjusted between different VNF instances.

[0083] As a preferred, in order to ensure that the resource joint scheduling plane can reflect the resource state of the computing node in real time, the NFV orchestrator adopts an adaptive resource mapping algorithm, which can automatically trigger a heterogeneous computing power adaptation strategy when the computing resource is insufficient, for example, when the CPU resource is overloaded, part of the computing task of the VNF instance can be automatically migrated to the DPU (data processing unit) or FPGA acceleration unit, thereby improving the computing efficiency. The resource joint scheduling plane not only can accurately match the computing resource and demand, but also can dynamically optimize the deployment of VNF instances in a super large scale computing environment, improve the resource utilization rate of the whole cloud computing platform, and ensure the stability and scalability of business computing.

[0084] The specific steps of S1.3 are as follows:

[0085] S1.3.1: mapping the physical resource indicators and the demand characteristics to obtain a resource mapping relationship, wherein the physical computing resource is associated with the computing resource demand, the storage resource is associated with the storage demand, and the network resource is associated with the network performance demand;

[0086] In this embodiment, in order to ensure the efficient matching of computing resources, storage resources and network resources, it is necessary to first establish the association mapping of physical resources and demand characteristics, so that the demand of the computing task can be accurately corresponded to the most suitable physical node.

[0087] Specifically, after the data collection is completed, it is matched one-to-one with the corresponding demand characteristics according to the resource category, wherein the computing resource mainly matches the computing-intensive task, the storage resource adapts the storage-intensive task, and the network resource adapts the high-throughput or low-latency demand service. The association mapping not only considers the current use of the resource, but also combines historical data analysis, so as to dynamically adjust the mapping relationship and ensure the coordinated optimization of computing, storage and network. For example, in the AI training task scenario, the demand of the computing task may change with the training progress, and in the initial stage, it may mainly depend on the computing resource, while in the later stage of training, the data access demand increases, and the storage resource demand may dominate. The traditional method often uses a fixed resource allocation method, which cannot dynamically adapt to such demand changes, while the present embodiment can adjust the resource matching method according to the task running state through the resource association mapping mechanism, thereby improving the resource utilization rate and reducing the problem of decreased computing efficiency caused by resource mismatch. In addition, the mapping method can also be applied to a multi-tenant environment. In a cloud computing platform, the business demands of different tenants are different, and the priorities of resource demand may also be different, so it is necessary to dynamically adjust the mapping strategy based on the weight calculation method to ensure that high-priority businesses can be allocated the best resources, while low-priority businesses can be allocated suboptimal resources in resource shortage, so as to realize fair scheduling and maximize resource utilization.

[0088] S1.3.2: According to the resource mapping relationship, a resource adaptation matrix is constructed through a predefined resource adaptation rule, wherein the resource adaptation matrix standardizes the resource states of each physical node to form a two-dimensional adaptation matrix of physical resource indicators and demand characteristics, and the resource adaptation matrix is used to describe the adaptation degree of different physical nodes in each demand dimension;

[0089] In this embodiment, in order to convert the resource mapping relationship into a calculable optimization model, it is necessary to construct a resource adaptation matrix, that is, to standardize the physical computing, storage and network resources in a two-dimensional space to quantify the adaptation degree of different resource nodes to various task demands. The construction of the resource adaptation matrix first relies on a predefined resource adaptation rule, which sets the standardization parameters of each resource dimension based on historical data analysis, task execution mode and physical resource load. For example, in terms of computing resource adaptation, the computing power units of CPU and GPU are different, and traditional resource allocation methods cannot directly compare the two. Therefore, this embodiment uses the floating point operation demand (FLOPs) of the computing task as a benchmark to convert the computing power of CPU and GPU into a unified FLOPs unit, so that they can be directly matched. In addition, for storage resources, traditional storage adaptation methods are mainly based on capacity matching, but this embodiment adds indexes such as I / O throughput capacity and cache hit rate in the adaptation matrix, so that storage-intensive tasks can be preferentially scheduled to storage nodes with low latency and high throughput to reduce data access bottlenecks. For network resources, the topology of the network link is considered, and the network latency, packet loss rate and bandwidth between computing nodes are used as the measurement standard, and an adjacent node priority matching strategy is constructed in the adaptation matrix, that is, high interaction VNF components are preferentially scheduled to physically adjacent computing nodes to reduce the additional overhead of cross-domain data transmission. After the construction of the adaptation matrix, the capabilities of different physical resource nodes are balanced through standardization calculation methods, for example, the computing capacity, storage capacity and network capacity of each physical node in the matrix are converted into normalized values in the range of 0-1, ensuring the calculation efficiency of the resource matching process.

[0090] Specifically, the resource adaptation matrix also involves dynamic adjustment of resource adaptation. Resource allocation is usually based on static rules, such as predefining that a certain task type is always allocated to a fixed computing node, which cannot be flexibly adjusted when the task load changes. This embodiment adopts a dynamic weight allocation mechanism to monitor the load changes of computing nodes in real time and dynamically adjust the weights in the adaptation matrix, for example, when the CPU utilization of a certain computing node approaches the threshold, its adaptation score will automatically decrease, so that it is preferentially excluded in the matching process, while the nodes with lower computing load are preferentially selected. Such dynamic adjustment method can ensure the load balancing of computing resources and avoid affecting the overall business performance due to overload of individual computing nodes.

[0091] Further, the resource adaptation matrix of the embodiment also supports adaptive strategy optimization, that is, based on machine learning analysis of historical task execution, automatically adjusting the adaptation rules. For example, in a cloud computing environment, the resource demand of different types of tasks may have periodic changes in different time periods, for example, an AI training task consumes a large amount of GPU during the daytime peak period, and the GPU demand decreases during the night low peak period, and a storage task may dominate. The embodiment can analyze these trends based on historical data and dynamically adjust the task mapping relationship in the resource adaptation matrix, for example, prioritizing computing task resources during peak periods and adjusting storage priority during low peak periods, thereby achieving higher resource utilization.

[0092] The specific steps of S2 are as follows:

[0093] S2.1: According to the resource joint scheduling plane, a VNF deployment optimization model based on hierarchical decomposition is constructed;

[0094] In the embodiment, the deployment optimization of VNF instances adopts a hierarchical decomposition strategy, and efficient allocation of computing resources, storage resources and network resources is performed through two levels of global resource domain planning and local node deployment optimization. First, in the resource joint scheduling plane, a VNF function demand mapping is constructed in a two-dimensional coordinate system, in which the horizontal axis represents the distribution of physical computing resources and the vertical axis represents the demand characteristics of VNF instances. The resource polygon and demand rectangle overlap degree calculation is used to determine the resource adaptation. In order to ensure the optimality of resource allocation, a hierarchical decision tree is constructed based on the mapping, wherein the first layer represents the global scheduling target of the VNF instance, the second layer is refined into sub-task allocation within the resource domain, and the third layer is further divided into specific physical nodes. The core idea of hierarchical decomposition is to reduce the computational complexity of the VNF deployment problem, and to decompose a large-scale task into multiple smaller decision problems, thereby reducing the difficulty of solving NP-hard problems.

[0095] As preferred, the present application allows filtering out non-adapted computing resources in the global planning stage, avoiding the computational resource redundancy caused by the traditional traversal matching method, and improving the computing efficiency. In addition, when constructing the optimization model, a resource load prediction mechanism is also introduced, the computing, storage and bandwidth utilization of each resource node is analyzed using historical task load data, and combined with the time series prediction method, the allocation strategy of the computing resources is adjusted in advance to ensure that the VNF instance deployment will not cause resource congestion due to sudden load. Compared with the traditional static scheduling method, the hierarchical decomposition strategy of the embodiment can realize fast calculation, accurate matching and dynamic adjustment in a super large-scale computing environment, ensuring that the VNF instance can be deployed efficiently and adapt to real-time changes in resource state.

[0096] S2.2: Determine the target resource domain of the VNF instance through global resource domain planning;

[0097] S2.3: matching VNF instances with physical computing, storage and network resources of physical nodes according to target resource domains through local node deployment optimization;

[0098] S2.4: updating the matching result through a dynamic computing power migration algorithm to generate a deployment decision;

[0099] In this embodiment, in order to further optimize the deployment of VNF instances and ensure dynamic balancing of computing resources, a dynamic computing power migration algorithm is adopted, that is, during task execution, computing tasks are dynamically migrated according to real-time load conditions of computing resources. The core idea of the algorithm is to calculate an optimal computing power migration path based on an A* search strategy, that is, in the resource joint scheduling plane, a target node with the lowest computing migration cost and the lowest network transmission cost is searched.

[0100] Specifically, first, the load condition of the current computing node is monitored, and if the CPU / GPU utilization of a certain computing node exceeds a preset threshold, the computing power migration demand is triggered. Then, based on the topological relationship of the resource joint scheduling plane, an optimal migration path is calculated, and the goal is to find a target node with optimal computing resources and the lowest network transmission cost, and perform task migration.

[0101] Further, in order to reduce the data synchronization overhead in the computing power migration process, the embodiment adopts an incremental computing snapshot mechanism, that is, during task migration, only the incremental data of the computing task is synchronized, rather than the complete migration computing state, thereby reducing the data transmission amount and improving the migration efficiency.

[0102] As a preferred, in order to improve the stability of task migration, the embodiment also introduces an adaptive load balancing strategy, which dynamically adjusts the resource joint scheduling plane after the computing power migration is completed, to ensure that the new computing resource layout conforms to the global optimal allocation strategy. Compared with the traditional static task scheduling method, the dynamic computing power migration algorithm of the embodiment can greatly improve the utilization rate of computing resources, reduce the load imbalance problem of computing nodes, and ensure the stable operation of VNF instances, improve the computing efficiency and resource utilization of the business.

[0103] The specific steps of S2.2 are as follows:

[0104] S2.2.1: establishing a plane coordinate graph of resources and demands in a two-dimensional coordinate system according to the resource joint scheduling plane, wherein the horizontal axis represents the physical resource computing power distribution and the vertical axis represents the demand characteristics;

[0105] In this embodiment, in order to realize efficient allocation of computing resources, first, a two-dimensional mapping relationship of resources and demands is established in the resource joint scheduling plane, that is, a plane coordinate graph is constructed with the horizontal axis representing the physical resource computing power distribution and the vertical axis representing the demand characteristics.

[0106] Specifically, the physical resource computing power distribution includes computing power, storage power and network power, while the demand characteristics cover computing demand, storage demand and network demand. Mapping resources and demand to this two-dimensional space enables the scheduling system to perform resource adaptation calculation under an intuitive spatial model, thereby reducing the complexity of resource scheduling and improving computing efficiency. Through this mapping relationship, the distribution state of resources can be quickly identified, and the overloaded and idle areas of computing power can be found, so that the task scheduling system can preferentially select computing nodes with low load for resource allocation, avoiding uneven resource utilization. At the same time, this two-dimensional coordinate mapping method can also be applied to a multi-tenant computing environment. In a cloud computing platform, the computing demands of different tenants may differ, and this method can ensure that the computing resource demands of different tenants can be independently managed and reasonably allocated according to their business priorities, improving the fairness and scalability of the overall computing cluster.

[0107] S2.2.2: mapping the physical nodes as resource polygons and the demand of the VNF instance as demand nodes, and mapping the demand nodes as demand rectangles;

[0108] In this embodiment, in order to make resource scheduling more intuitive and calculable, a geometric mapping method is used to map physical computing resources as resource polygons and computing demand as demand rectangles. The construction of resource polygons is based on the actual computing power, storage capacity and network bandwidth of physical nodes. A resource utilization-based shape analysis method is used to represent high-power nodes as larger polygons and low-power nodes as smaller polygons. Meanwhile, considering the proportion of storage and network resources, different shapes of resource polygons are used to represent the heterogeneous characteristics of different computing resources. The construction of demand rectangles is based on the computing demand, data transmission demand and storage demand of VNF instances. A normalized demand modeling method is used to standardize the demand of different computing tasks, ensuring that it can be matched with resource polygons on a two-dimensional plane. This method can provide a geometric optimization calculation method. Compared with the traditional table query or rule matching method, the geometric mapping method can directly determine the resource adaptation by calculating the intersection area of the resource polygon and the demand rectangle, improving the matching efficiency and reducing the computational complexity of the computing resource scheduling. When the computing task changes, the size of the demand rectangle can be adjusted in real time to simulate the trend of the computing task, thereby dynamically adjusting the matching relationship of the computing resources and improving the adaptability of the scheduling system.

[0109] S2.2.3: calculating the overlap degree of the resource polygon and the demand rectangle, and taking the resource polygon with an overlap degree greater than an overlap threshold as a target resource domain of the corresponding demand node;

[0110] In this embodiment, the key of resource matching lies in the overlap calculation of resource polygons and demand rectangles, and the core idea is to determine the optimal computing resource matching scheme by calculating the intersection area between resource supply and demand.

[0111] Specifically, the overlap calculation adopts a weighted geometric matching method, that is, different weights are used to calculate the resource adaptation degree considering the priority of different computing tasks. For example, the weight of CPU / GPU computing power is higher when matching the computing-intensive tasks, and the weight of storage I / O throughput is higher for storage-intensive tasks, so that different types of tasks can find the most suitable computing nodes for deployment. If the calculated overlap is greater than the set overlap threshold, the resource polygon is determined as the target resource domain of the demand rectangle, and is used as the candidate set of computing nodes for VNF instances.

[0112] Further, in order to improve the calculation efficiency, the embodiment introduces a fast screening mechanism, which preferentially screens the computing nodes that may meet the matching requirements when calculating the overlap, and only performs detailed geometric matching calculation on these nodes, so as to reduce unnecessary calculation overhead and improve the calculation speed.

[0113] Further, the present application can also be applied to dynamic resource expansion. When the load of the computing task increases, the coverage of the resource polygon can be automatically adjusted to ensure that the computing task always has sufficient computing resource support and improve the stability of task execution and resource utilization.

[0114] Please refer to Figure 2 The principle diagram of target resource domain selection of the embodiment of the present application is shown in FIG. 1, wherein A, B and C represent resource polygons respectively, and D represents a demand rectangle. The matching relationship between resources and demand is intuitively displayed on a two-dimensional plane. The implicit resource conflicts (such as insufficient computing power and mismatched network performance) in the traditional scheme are explicitly mapped as geometric intersection areas, so that the problems can be quickly identified. The large intersection area between the demand rectangle D and the resource polygon A directly reflects the optimal adaptability between D and A.

[0115] S2.2.4: Split each heterogeneous VNF into multiple functional components, and define the synchronization communication frequency between the components, wherein each functional component corresponds to different demand characteristics;

[0116] In this embodiment, in order to improve the flexibility of computing resources, the VNF instance is split into multiple independent functional components, and the synchronization communication frequency is defined according to the data interaction demand between the components. The split is mainly based on computing-intensive tasks, storage-intensive tasks and network-intensive tasks. In this way, each component can be independently matched to the most suitable computing resource. For example, a complex VNF instance can include a data preprocessing module (computing-intensive), a database storage module (storage-intensive) and a real-time inference module (network-intensive). In the traditional method, these functional components are usually uniformly scheduled to a single computing node, resulting in unbalanced utilization of computing resources. In this embodiment, they can be respectively scheduled to the most suitable computing resource, improving the overall efficiency of task execution. At the same time, in order to reduce the data transmission overhead, this embodiment introduces a dynamic synchronization mechanism to adjust the data synchronization strategy according to the interaction frequency between components, so as to ensure that components with high-frequency interaction can be deployed in a low-latency computing environment, while components with low-frequency interaction can prioritize the utilization of computing resources, thereby achieving comprehensive optimization of high throughput and low latency.

[0117] S2.2.5: Within each target resource domain, the functional components are assigned to the demand node with the highest similarity to their demand characteristics, wherein the similarity is calculated by cosine similarity;

[0118] In this embodiment, in order to ensure accurate allocation of computing resources, a cosine similarity calculation method is used. In the matching process, based on the vectorized characteristics of computing resources, storage resources and network resources, the similarity score between the demand node and the functional component demand is calculated, and the node with the highest similarity is selected for resource matching.

[0119] Further, in the case where a single computing node cannot meet the component demand, a cross-node collaborative computing method is used, that is, through a distributed task scheduling system, the computing task is split into multiple parallel computing units and deployed to different computing nodes, and a detailed network flow table is automatically issued by the SDN controller to ensure the collaborative execution efficiency of the computing task among multiple nodes. Compared with the traditional centralized computing mode, this method can better utilize heterogeneous computing resources, improve the scalability of computing tasks, and ensure that the computing tasks of the VNF instance can be efficiently executed in a super-large-scale computing environment, improve the task throughput, and at the same time reduce the problem of resource fragmentation.

[0120] S2.2.6: If a single node cannot meet all the resource demands of the component, the functional component is split into sub-components that can run across nodes in collaboration, and a detailed network flow table is issued by the SDN controller;

[0121] The specific steps of S2.3 are as follows:

[0122] 2.3.1: constructing a running container on each demand node, wherein the running container comprises an initial container and a standby container, the initial container is used to carry a currently executing functional component, and the standby container is in an idle state;

[0123] In this embodiment, in order to improve the deployment flexibility of the VNF instance and ensure that the computing resources can be quickly adapted to task scheduling in the case of burst load, a running container is constructed on each demand node, which includes an initial container and a standby container. The initial container is a computing unit that actually carries the currently running functional component, and its computing resources, storage resources and network bandwidth have been occupied by the current task. The standby container is a resource reservation unit in an idle state, which does not actively undertake task computing, but has the same resource configuration as the initial container, can be activated immediately when the computing demand increases, and avoids the decline of business performance caused by the delay of computing resource allocation. The construction of the running container uses dynamic instantiation technology, that is, when the computing node is initialized, not only the initial container currently required is created, but also the future possible resource demand is predicted based on the data statistics of the historical computing tasks, and the corresponding standby container is created according to the set proportion, so that it can be switched to the execution state at any time. The resource application delay when starting the task can be reduced, the computing task can complete the computing resource adaptation in milliseconds, and it is suitable for high-concurrency and high-computing-load cloud computing environment.

[0124] Further, the existence of the standby container can also effectively cope with the sudden change of computing resource demand. For example, in AI inference tasks or high-performance computing tasks, the computing load may increase dramatically in a short time, and the traditional static resource allocation strategy is difficult to complete resource scheduling in a short time. However, the standby container mechanism of the present embodiment can complete resource expansion in a short time, ensure the smooth execution of the computing task, and improve the computing throughput of the system.

[0125] 2.3.2: judging whether the demand node has a secondary overlap state according to the running state of the running container, wherein the secondary overlap state refers to the existence of competition conflict between demand nodes;

[0126] In this embodiment, in order to ensure the stability of the computing task and avoid the execution delay or failure of the computing task caused by resource competition, it is necessary to monitor the computing resource usage of the demand node in real time and judge whether there is a secondary overlap state. The secondary overlap state refers to the competition conflict of resource demand of multiple demand nodes, that is, the computing tasks of different VNF instances compete for the same computing, storage or network resources due to uneven resource allocation during the running process. For example, in a GPU computing environment, if multiple tasks are scheduled to the same GPU at the same time, the computing resource competition may be caused due to the computing load exceeding the concurrent processing capacity of the GPU, which affects the task execution efficiency.

[0127] Specifically, the embodiment monitors the CPU usage, GPU utilization, I / O throughput and network bandwidth occupation of the running container in real time through an intelligent monitoring mechanism, and analyzes the usage trend of the current resources using a load balancing prediction algorithm. If the resource utilization of a demand node exceeds the set threshold, and the resource usage of its adjacent demand nodes is also in a high load state, it is determined that there is a secondary overlap state in the current environment. The detection of the secondary overlap state is based on a multi-dimensional resource state evaluation model, that is, not only the CPU, memory and storage load of the current computing node are considered, but also a network topology awareness mechanism is introduced to analyze the network transmission of the computing task, so as to ensure that the use efficiency of computing resources and network resources can be optimized at the same time when scheduling the computing task. The advantage of this real-time monitoring mechanism is that it can discover potential resource competition problems before the computing resources are tight, and take optimization measures in advance to avoid system instability problems caused by uneven allocation of computing resources, and improve the overall computing efficiency of VNF instances.

[0128] 2.3.3: When there is a secondary overlap state, the secondary overlap part is redistributed;

[0129] In this embodiment, when it is detected that there is a secondary overlap state between demand nodes, the computing tasks in the secondary overlap part need to be redistributed to ensure the fairness and efficient use of computing resources. The core idea of redistribution is based on a real-time load adjustment mechanism, that is, by using an intelligent computing task scheduling strategy, the resource demand of the computing task is redistributed to the computing nodes with low load.

[0130] Specifically, a dynamic load balancing algorithm is used to migrate the tasks with high resource demand according to the load of the current computing task, and preferentially select the nodes with idle or low load computing resources as target nodes. In addition, in order to avoid resource waste caused by frequent migration of computing tasks, a task migration prediction model is introduced in this embodiment to evaluate the resource adaptability before task migration, to determine whether the target node has the ability to receive the migrated task, and to calculate the migration cost to ensure that the computing task after migration will not affect the overall task execution efficiency due to data synchronization problems or unreasonable computing load allocation. In the redistribution process, an adaptive resource scheduling strategy is used to process different types of computing tasks according to their priorities, for example, for high-priority real-time computing tasks, the stability of their computing resources is preferentially guaranteed, while for low-priority background computing tasks, they are allowed to be delayed or migrated to nodes with weak computing ability when resources are tight, so as to ensure that the overall computing resources of the system can achieve optimal utilization. This can effectively alleviate the competition problem of computing resources, improve the execution stability of computing tasks, and ensure the efficient use of computing resources.

[0131] The specific steps of S2.3.3 are as follows:

[0132] S2.3.3.1: calculating a conflict factor according to a change trend of a corresponding demand rectangle in a resource joint scheduling plane;

[0133] In this embodiment, in order to dynamically monitor the competition of computing resources, a conflict factor needs to be calculated to quantify the degree of competition of computing resources between different demand nodes, so as to take corresponding adjustment strategies when resource conflicts occur.

[0134] Specifically, the calculation of the conflict factor is based on a resource joint scheduling plane, in which each demand rectangle represents the computing resource demand of a VNF instance, and the resource polygon represents the computing power supply of a physical node. When a demand rectangle deforms, such as area increase (indicating that the task computing demand rises) or position offset (indicating that the task migrates between different computing nodes), it may indicate that the task is experiencing resource competition.

[0135] Further, this embodiment adopts a time sequence resource utilization monitoring mechanism, that is, by monitoring the change trend (such as deformation rate, increase or decrease amplitude of covering the resource polygon) of the demand rectangle in a period of time, the matching degree of resource supply and demand is calculated, and the current load state of the computing resource is combined to comprehensively calculate the conflict factor. The conflict factor reflects the resource competition intensity between a demand node and adjacent demand nodes. If the conflict factor of a demand node continuously rises, it indicates that the resource demand of the node has exceeded the carrying capacity of the resource polygon, and optimization adjustment needs to be performed. Preventing the decrease of computing task execution efficiency due to resource contention, improving the foresight of resource scheduling, thereby reducing the load imbalance phenomenon of VNF instances in a super large-scale computing environment.

[0136] S2.3.3.2: if the conflict factor is less than a preset conflict threshold, the running container remains the current running state;

[0137] In this embodiment, if the calculated conflict factor is less than the set conflict threshold, it indicates that the resource competition situation of the current demand node has not reached a degree that seriously affects the execution of the computing task, so resource adjustment is not needed, and the running container can remain in the current computing state. In this case, although there is a certain degree of resource competition, the system can still be adapted through normal computing resource scheduling strategies, without the need for large-scale task migration or container expansion. Avoiding over-scheduling, preventing additional computing overhead due to frequent resource adjustment, especially in the case of large fluctuations of computing tasks in a short period of time, over-scheduling may cause resource waste.

[0138] S2.3.3.3: if the conflict factor is greater than or equal to the preset conflict threshold, allocating the prepared container to the demand node in the secondary overlapping state;

[0139] In the present embodiment, when the conflict factor exceeds the set conflict threshold value, it indicates that the resource demand of the current computing task has exceeded the carrying capacity of a single computing node, and a secondary overlap state has occurred, i.e., different demand nodes compete for the same computing, storage or network resources. In this case, computing resource expansion is needed, i.e., the standby container is activated and allocated to the demand node with the most serious resource competition. The activation mode of the standby container adopts an automatic instantiation mechanism based on resource priority, i.e., the demand node with the heaviest resource load and the highest conflict factor is preferentially selected for expansion, and the computing resource configuration of the standby container is ensured to be consistent with that of the initial container to realize seamless migration of the task. Dynamic allocation of the standby container can significantly reduce the resource competition of the computing task, improve the throughput of the computing task, and reduce the delay problem of task execution.

[0140] Further, in allocating the standby container, the present embodiment also adopts a load balancing scheduling strategy, i.e., in the case where resource conflicts exist in multiple demand nodes, the resource allocation priority is sorted according to the conflict factor, ensuring that the node with the most resource tension can first get resource supplement, thereby optimizing the overall utilization efficiency of computing resources and avoiding the problem of uneven resource allocation.

[0141] S2.3.3.4: When the standby container is insufficient to meet the demand of the demand node in the secondary overlap state, the demand rectangle splitting mechanism is used to split the functional components currently in conflict on each demand node into multiple independent sub-tasks, and the re-allocation is performed according to the priority;

[0142] In the present embodiment, when all available standby containers have been allocated and still cannot meet the resource demand of the demand node, the demand rectangle splitting mechanism needs to be used, i.e., the computing task with the most serious resource conflict is split into multiple independent sub-tasks, and the re-scheduling is performed according to the execution priority of the task. The core idea of demand rectangle splitting is to split the computing task according to the functional components, so that it can be executed in parallel on different computing nodes, thereby reducing the load pressure of a single computing node.

[0143] Specifically, the task splitting method includes: hierarchical splitting based on computing dependency, i.e., modularizing the computing task according to data flow and computing process, splitting into multiple computing units that can be independently executed, and according to the execution priority of the computing unit, preferentially scheduling the most critical task to the node with the most sufficient computing resource; task segmentation based on data parallelism, i.e., for tasks that can be parallel computed, such as deep learning inference tasks or large-scale data processing tasks, by dividing data blocks, multiple computing nodes simultaneously process different data subsets to improve computing throughput. The advantage of the demand rectangle splitting mechanism is that even in the case of extremely tight computing resources, the computing stability and task execution efficiency of the system can be improved through the reconstruction and allocation optimization of computing tasks, and the computing task will not be in a waiting state for a long time due to resource competition problems.

[0144] Further, after task splitting, a refined network flow table is issued by the SDN controller to ensure that the network communication between the split sub-tasks is optimal, avoiding the decline of computing performance caused by data synchronization delay. Compared with the traditional computing resource scheduling method, the demand rectangle splitting mechanism of the embodiment can greatly improve the computing efficiency of the VNF instance, reduce the task execution time, improve the utilization rate of computing resources, and enhance the adaptability of the system in a high-concurrency computing environment.

[0145] Embodiment 2

[0146] Please refer to Figure 3 The present application provides an embodiment: an enterprise-level super-large networking system based on SDN and NFV, which comprises a data acquisition module, a plane construction module and an instance deployment module.

[0147] The data acquisition module is used to acquire physical resource indicators and demand description files.

[0148] The plane construction module is used to process the physical resource indicators and demand description files through the SDN and the NFV, and construct a resource joint scheduling plane.

[0149] The instance deployment module is used to match the VNF instance according to the resource joint scheduling plane, and generate a deployment decision of the VNF instance.

[0150] The instance deployment module comprises:

[0151] A global planning unit is used to determine the target resource domain of the VNF instance.

[0152] A local optimization unit is used to match the physical computing, storage and network resources of the VNF instance and the physical node according to the target resource domain.

[0153] The matching updating unit is configured to update the matching result by using a dynamic computing power migration algorithm to generate a deployment decision.

[0154] Although the embodiments of the present application have been shown and described above, it should be understood by those skilled in the art that the above embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for enterprise-level super-large networking based on SDN and NFV, applied to a network service component of a cloud computing platform, wherein the super-large networking needs to satisfy that the cloud computing platform manages greater than or equal to 10,000 virtualized network devices and concurrently runs greater than or equal to 1,000 heterogeneous VNF instances, and the method is characterized in that, The enterprise-level super-large networking method comprises: building a resource joint scheduling plane through the SDN and the NFV; matching the heterogeneous VNF instances according to the resource joint scheduling plane to generate a deployment decision of the heterogeneous VNF instances; deploying a distributed probe cluster in the network service component to update the deployment decision in real time to perform super-large networking on the enterprise, wherein the probe cluster is used to monitor running data of the VNF instances; matching the heterogeneous VNF instances to generate a deployment decision of the heterogeneous VNF instances, comprising: building a VNF deployment optimization model based on hierarchical decomposition according to the resource joint scheduling plane to generate a matching result, wherein the VNF deployment optimization model comprises global resource domain planning and local node deployment optimization, the global resource domain planning is used to determine a target resource domain of a VNF instance, and the local node deployment optimization is used to match the VNF instance with physical computing, storage and network resources of a physical node according to the target resource domain; updating the matching result through a dynamic computing power migration algorithm to generate a deployment decision; the global resource domain planning comprises: establishing a plane coordinate graph of resources and demands in a two-dimensional coordinate system according to the resource joint scheduling plane, wherein a horizontal axis represents physical resource computing power distribution and a vertical axis represents demand characteristics; mapping a physical node as a resource polygon and mapping demand of a VNF instance as a demand node, and mapping the demand node as a demand rectangle; calculating an overlap degree of the resource polygon and the demand rectangle, and taking the resource polygon with the overlap degree greater than an overlap threshold as a target resource domain of the corresponding demand node; splitting each heterogeneous VNF into multiple functional components, and defining a synchronous communication frequency among the components, wherein each functional component corresponds to different demand characteristics; allocating the functional components to demand nodes with the highest similarity in demand characteristics in each target resource domain, wherein the similarity is calculated by a cosine similarity; if a single node cannot meet all resource demands of a component, splitting the functional component into sub-components that can be cooperatively operated across nodes, and issuing a refined network flow table by an SDN controller; updating the matching result through a dynamic computing power migration algorithm, comprising: calculating a load balancing degree of a current resource polygon according to the matching result; if the load balancing degree is greater than a preset load threshold, triggering a computing power migration demand; calculating a computing power migration factor of a target resource domain where the computing power migration demand occurs, wherein the computing power migration factor comprises a computing power overflow degree, a data transmission cost and a migration time estimation; matching the computing power migration factor through an A* algorithm according to the resource joint scheduling plane to obtain a second target domain, wherein the A* algorithm calculates a migration path according to a topology of a remaining idle resource domain of the resource joint scheduling plane, and selects an idle resource domain with the minimum calculation cost; migrating to the second target domain, and updating the resource joint scheduling plane according to a load status of a target node after the computing power migration, and recalculating a matching relationship between the resource polygon and the demand rectangle to generate a deployment decision. 2.The enterprise-level super-large networking method based on SDN and NFV according to claim 1, characterized in that, Building a resource joint scheduling plane through the SDN and the NFV, comprising: The SDN controller collects physical resource indexes of the virtualized network device based on an extended OpenFlow protocol, wherein the physical resource indexes comprise physical computing resources, storage resources and network resources, and stores the physical resource indexes in a resource state database; The NFV orchestrator analyzes a requirement description file submitted by an enterprise user, and analyzes requirement characteristics, wherein the requirement characteristics comprise computing resource requirements, storage requirements and network performance requirements; The physical resource indexes and the requirement characteristics are interfaced through a southbound interface with a KVM virtualization layer of a cloud platform to generate the resource joint scheduling plane. 3.The enterprise-level super-large networking method based on SDN and NFV according to claim 2, characterized in that, The interfacing of the physical resource indexes and the requirement characteristics comprises: The physical resource indexes and the requirement characteristics are associated and mapped to obtain a resource mapping relationship, wherein the physical computing resources are associated with the computing resource requirements, the storage resources are associated with the storage requirements, and the network resources are associated with the network performance requirements; According to the resource mapping relationship, a resource adaptation matrix is built through predefined resource adaptation rules, wherein the resource adaptation matrix standardizes resource states of each physical node to form a two-dimensional adaptation matrix of the physical resource indexes and the requirement characteristics, and the resource adaptation matrix is used to describe adaptation degrees of different physical nodes in each requirement dimension. 4.The enterprise-level super-large network method based on SDN and NFV of claim 1, wherein, The local node deployment optimization comprises: On each requirement node, a running container is built, wherein the running container comprises an initial container and a standby container, the initial container is used to carry a functional component that is currently being executed, and the standby container is in an idle state; According to a running condition of the running container, it is judged whether the requirement node has a secondary overlap state, wherein the secondary overlap state refers to a competitive conflict between requirement nodes; When the secondary overlap state exists, the secondary overlap part is reassigned.

5. The enterprise-level super-large network method based on SDN and NFV according to claim 4, characterized in that, The reassignment of the secondary overlap part comprises: A conflict factor is calculated according to a change trend of a corresponding requirement rectangle in the resource joint scheduling plane; If the conflict factor is less than a preset conflict threshold, the running container remains in a current running condition; If the conflict factor is greater than or equal to the preset conflict threshold, the standby container is assigned to the requirement node in which the secondary overlap state exists; When the standby container is insufficient to meet the requirement of the requirement node in which the secondary overlap state exists, a requirement rectangle splitting mechanism is used to split the functional component that currently has a conflict into multiple independent subtasks on each requirement node, and the subtasks are reassigned according to priorities.

6. An enterprise-level super-large networking system based on SDN and NFV, used for implementing the enterprise-level super-large networking method based on SDN and NFV according to any one of claims 1-5, characterized in that, The system comprises a data acquisition module, a plane building module and an instance deployment module; The data acquisition module is used to acquire physical resource indexes and requirement description files; The plane building module is used to process the physical resource indexes and the requirement description files through the SDN and the NFV to build a resource joint scheduling plane; The instance deployment module is used to match VNF instances according to the resource joint scheduling plane to generate deployment decisions of the VNF instances. 7.The enterprise-level super-large network system based on SDN and NFV according to claim 6, characterized in that, The instance deployment module comprises: A global planning unit configured to determine target resource domains for the VNF instances; A local optimization unit configured to match the VNF instances with physical computing, storage and network resources of the physical nodes according to the target resource domains; A matching updating unit configured to update the matching results by a dynamic computing power migration algorithm to generate a deployment decision.

Images