Three-layer decoupling scheduling system and method for distributed overlay network, medium, program product and terminal

Abstract

The application discloses a three-layer decoupling scheduling system and method for a distributed overlay network, a medium, a program product and a terminal, and relates to the technical field of network scheduling. The system comprises a semantic layer, a control bus and a data bus, and the three are completely decoupled in responsibility. The semantic layer adopts a minimum semantic input model, integrates tenant identification, service priority, node capability and link index into standardized unified routing decision input. The control bus is responsible for state acquisition and static image capture, and does not intervene in data plane forwarding. The data bus only performs data packet encapsulation forwarding and lightweight link observation, and does not participate in policy judgment. The application is based on the separation of the control plane and the data plane and the three-layer decoupling architecture, solves the problems of architecture coupling, semantic fragmentation, high expansion risk and lack of standardized feedback loop in the prior art, improves the stability, accuracy and evolvability of distributed overlay network scheduling, and realizes adaptive optimization of routing scheduling and a policy loop without blocking the hot path of the data plane.

Description

Technical Field

[0001] This invention belongs to the field of distributed overlay networks and software-defined networks, specifically involving a three-layer decoupled scheduling system, method, medium, program product, and terminal for distributed overlay networks. Based on a control plane and data plane separation architecture, it achieves layered decoupling, adaptive decision-making, and policy closed-loop optimization of routing scheduling in distributed overlay networks through static policy interfaces, a unified semantic input model, and an asynchronous feedback event mechanism. Background Technology

[0002] In current distributed overlay networks, virtual private networks, edge relay networks, and P2P / Relay hybrid network architectures, routing scheduling and data forwarding systems generally suffer from core technical defects such as severe architectural coupling, semantic fragmentation, high scalability risks, and lack of policy closure loops. These defects are specifically manifested in the following aspects:

[0003] First, there is a drifting coupling between control logic and data forwarding responsibilities. The existing network architecture mixes control logic such as policy judgment, permission auditing, state update, and fault handling with data plane hot path logic such as packet encapsulation, forwarding, and transmission. This results in high redundancy in the data plane forwarding process, which can easily cause problems such as forwarding delay jitter and instantaneous traffic congestion. Moreover, a single module failure can easily spread to the entire forwarding link, significantly reducing the overall stability of the network.

[0004] Second, business semantic fields are severely fragmented. In existing network systems, core scheduling semantic fields such as tenant identifiers, namespaces, service types, service priorities, and node service capabilities are stored and processed in different functional modules, lacking a unified structured input carrier. The routing decision-making process lacks standardized semantic basis, resulting in poor adaptability of scheduling strategies and the inability to achieve fine-grained and differentiated scheduling for different services and tenants.

[0005] Third, dynamic expansion mechanisms pose high security and stability risks. To adapt to diverse business scheduling needs, existing solutions often introduce mechanisms such as dynamic expansion registration, online function loading, and dynamic interface compatibility binding in the data plane hot path. However, as the core forwarding link, dynamic changes to the data plane can easily lead to interface compatibility failures, permission vulnerabilities, and process anomalies, significantly increasing the security risks and stability hazards of network operation.

[0006] Fourth, there is a lack of a standardized model for the routing scheduling feedback loop. Existing networks can only simply record the success or failure of routing forwarding. For scenarios such as path failure, link fallback, path recovery, and fluctuations in link health status, there is a lack of unified and standardized event definitions, storage, and transmission mechanisms. This makes it impossible to provide complete data support for subsequent policy optimization, link auditing, and fault tracing, and it is difficult to form an adaptive scheduling loop.

[0007] Fifth, the boundaries of architectural capabilities are blurred, leading to high delivery and review risks. The existing network architecture mixes long-term iterative capabilities such as shared memory bus, complete user-space virtual stack, and dynamic expansion operation mechanism with core baseline capabilities such as basic forwarding and routing scheduling. This premature integration into the infrastructure system results in a blurred scope of product delivery and increases compliance risks in patent examination and technology implementation.

[0008] In summary, the existing distributed overlay network scheduling architecture suffers from problems such as high coupling, low scheduling accuracy, high expansion risk, lack of standardized adaptive closed loop, and ambiguous architectural capability boundaries. It cannot balance the stability of basic forwarding with the flexibility of service expansion. There is an urgent need for a three-layer decoupled adaptive scheduling technology solution that features layered decoupling, static expansion, asynchronous closed loop, and clear boundaries. Summary of the Invention

[0009] The purpose of this invention is to address the technical problems in existing distributed overlay network scheduling architectures, such as coupling of control and data logic, fragmented business semantics, high risks of dynamic expansion, lack of standardized feedback loops, and ambiguous architectural capability boundaries. It provides a three-layer decoupled scheduling system, method, medium, program product, and terminal for distributed overlay networks. This system achieves policy input and scheduling loops without blocking hot paths in the data plane, unifies business semantic decision-making criteria, reduces architectural expansion risks, and enables adaptive optimization of routing scheduling. Existing solutions couple dynamic expansion with forwarding hot paths, leading to state drift. This invention constructs non-blocking static decision anchors in the control plane through a 'static policy interface + unified routing decision input', and then drives semantic layer updates without interfering with forwarding through 'asynchronous feedback events'. The synergy of these three elements fundamentally resolves the contradictions between hot path security, semantic fragmentation, and adaptive loops.

[0010] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows:

[0011] A three-layer decoupled scheduling system for distributed overlay networks includes a data bus, a control bus, and a semantic layer. The data bus is responsible for packet carrying, encapsulation, decapsulation, forwarding execution, and lightweight link state observation. This lightweight link state observation includes, but is not limited to, one or more indicators based on latency, bit error rate or packet loss rate, and connection stability. It only undertakes data plane forwarding responsibilities and does not participate in control plane state management or policy judgment. The control bus is responsible for node registration, heartbeat detection, node discovery, session maintenance, route hints, capability negotiation, and control event output. The control bus also integrates a static image capture interface, which uses a preset lock-free circular buffer or static shared memory area to capture and output routing decision results and associated service semantic information non-blockingly. The semantic layer is an independent semantic processing layer that uses a minimal semantic input model to convert tenant identifiers, namespaces, service identifiers, service priorities, and node capability summaries into unified routing decision inputs. The system completes service policy matching through a static policy interface and achieves adaptive closed-loop optimization of routing policies through asynchronous feedback events.

[0012] Furthermore, the data bus includes at least one of an overlay network transmission adapter and an overlay network encapsulation data plane bearer module, and the data plane bearer module does not assume control plane state management responsibilities; the lightweight link state observation includes at least RTT latency detection, packet loss rate statistics, and path failure count statistics, providing link indicator support for routing decisions.

[0013] Furthermore, the adaptation methods supported by the overlay network transmission adapter include one or more of the following: tunnel-based transmission mechanism, point-to-point (P2P) direct communication mode, or relay forwarding mode; wherein, the tunnel-based transmission mechanism is preferably a TUN / TAP interface.

[0014] Furthermore, the unified routing decision input output by the semantic layer is structured data, including routing status information, relay candidate information, service semantic fields, node capability summaries, and link health indicators. The routing status information includes the current preferred route type, the current active route type, and the maximum hop limit. The relay candidate information includes the primary relay candidate node and the set of backup relay candidate nodes. The service semantic fields are used to characterize the service attributes of data transmission, including but not limited to virtual network identifiers, tenant identifiers, namespaces, service identifiers, and service priorities. The node capability summaries include whether P2P priority is supported, whether there are valid relay candidates, and the number of healthy relay nodes. The link health indicators include real-time transmission latency, link loss rate, and tunnel degradation status.

[0015] Furthermore, the static policy interface is implemented through compile-time preset assembly, configuration file-driven, or pre-set policy tables, and only supports static policy matching and business profile mapping. It does not include runtime plugin discovery, online hot loading, or dynamic ABI-compatible binding mechanisms. The service profiles that the static policy interface can map include default services, low-latency services, high-volume transmission services, P2P priority scheduling, Relay priority scheduling, and aggressive link recovery.

[0016] Furthermore, the asynchronous feedback event includes at least one of the following: P2P transmission failure, P2P transmission success, Relay transmission success, Relay path rollback, priority path recovery, cooling-off period activation, and path health update event. The asynchronous feedback event is bound to the corresponding routing decision result and carries a timestamp. It is output through an asynchronous bypass channel and can be asynchronously consumed by external modules without blocking or modifying the original data plane forwarding data stream. It is used to drive path rollback, cooling-off, recovery adaptive logic, and audit analysis.

[0017] Furthermore, the external module consumes the asynchronous feedback event non-blockingly through an asynchronous bypass channel or a local lightweight interface, and the consumption process does not modify or intercept the original data plane forwarding stream.

[0018] Furthermore, the system also includes access constraint rules, which are used to delineate the boundary between basic regression baseline capabilities and long-term post-extension capabilities. These access constraint rules are configured with a routing selection unit interception strategy. When a data stream containing dynamic extension registration instructions or runtime plugin loading requests is detected, the routing selection unit intercepts the request and directs it to an independent module for execution, preventing resource preemption of the data plane forwarding thread. The basic regression baseline capabilities include closed-loop scheduling based on relay forwarding mode, priority scheduling for point-to-point direct communication mode, and routing rollback and recovery mechanisms in case of node anomalies, link detection / cooling / recovery, and overlay network encapsulation metadata processing, constituting a stable and regressible delivery baseline for the system. The long-term post-extension capabilities include a dynamic extension operation mechanism, a shared memory bus, a complete user-space virtual stack, and a global strong consistency synchronization mechanism; these are iterative extension capabilities rather than essential basic capabilities. The long-term post-extension capabilities can only be accessed through static configuration or independent modules, and must reserve degradation and rollback paths. Access to data plane hot paths is prohibited, and bypassing the routing selection unit to customize extension scheduling logic is prohibited.

[0019] A data scheduling method, applied to the aforementioned three-layer decoupled scheduling system for distributed overlay networks, includes the following steps:

[0020] S1. Control plane status acquisition: Continuously acquire status metadata and routing basic metadata of all network nodes through the control bus;

[0021] S2. Unified Semantic Processing: The semantic layer extracts the business semantic fields from the metadata output by the control bus, combines the node capability summary and link health indicators, and generates standardized unified routing decision input.

[0022] S3. Static policy matching: The unified routing decision input is matched through the static policy interface to generate a determined target forwarding path;

[0023] S4. Data plane forwarding execution: The data bus completes packet encapsulation, decapsulation and link forwarding according to the target forwarding path, synchronously updates the lightweight link status observation indicators, and triggers the generation of asynchronous feedback events;

[0024] S5. Asynchronous Feedback Event Output: After binding a timestamp to the generated standardized asynchronous feedback event, it is output to the external module for consumption through a non-blocking bypass channel;

[0025] S6. Closed-loop adaptive optimization: The external module consumes the asynchronous feedback events, updates the path health status, and drives the adaptive iteration of the routing scheduling strategy.

[0026] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data scheduling method described above for a three-layer decoupled scheduling system for distributed overlay networks.

[0027] A computer program product, including computer program code, enables the terminal to implement the aforementioned data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks when the computer program product is run on a terminal.

[0028] An electronic terminal includes a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it implements the data scheduling method described above for a three-layer decoupled scheduling system for distributed overlay networks.

[0029] The advantages of this invention compared to the prior art are:

[0030] 1. Reduce the operational risks of data plane hot paths. This invention adopts an adaptive three-layer decoupled architecture, which completely isolates the control logic such as policy judgment, audit analysis, and state update from the data forwarding hot path. It replaces the traditional dynamic expansion mechanism with a static policy interface and achieves feedback closed loop through asynchronous bypass, avoiding control logic blockage and interference with data plane forwarding. This effectively solves the problems of forwarding latency jitter and fault propagation, and significantly improves network forwarding stability.

[0031] 2. Achieve standardized decision-making based on business semantics. This invention uses an independent semantic layer to unify fragmented fields such as tenants, services, priorities, node capabilities, and link status into standardized routing decision inputs. This enables refined and differentiated adaptation of routing scheduling, solving the problems of semantic confusion and inconsistent decision-making basis in traditional scheduling, and improving scheduling accuracy.

[0032] 3. Enhance system evolvability and scalability. This invention constructs a progressively scalable architecture through static interfaces and access constraint rules. Subsequent extension capabilities such as audit governance, QoS optimization, policy arbitration, and AI intelligent parameter tuning can all be accessed through static configuration and independent module mounting, without modifying the core forwarding architecture. It is compatible with existing baseline capabilities, and the extension process is secure and controllable.

[0033] 4. Enhance network auditability and traceability. This invention constructs a standardized routing feedback event and control event model, which fully records the entire process data of routing decisions, forwarding results, and link state changes. This provides complete and standardized data support for fault tracing, policy optimization, and compliance auditing, solving the problem of traditional network scheduling lacking a unified audit basis.

[0034] 5. Clearly define architectural capability boundaries to reduce delivery and review risks. This invention strictly distinguishes between basic baseline capabilities and long-term extended capabilities through access constraint rules, avoiding the premature inclusion of immature capabilities in the core architecture, preventing over-commitment of capabilities, and effectively reducing delivery, review, and online operation risks. Attached Figure Description

[0035] Figure 1 This is a diagram of the overall architecture of the three-layer decoupling scheduling system for distributed overlay networks according to the present invention.

[0036] Figure 2 This is a flowchart of the data scheduling method of the present invention.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] Example 1, in conjunction with Appendix Figure 1-2This paper presents a three-layer decoupled scheduling system for distributed overlay networks. It comprises a semantic layer (110), a control bus (120), and a data bus (130)—three independent layers, complemented by an asynchronous feedback bypass (140) and a static policy interface. Each layer's responsibilities are completely decoupled, with unidirectional data flow between layers. There is no functional overlap, logical coupling, or responsibility drift, fully meeting the core design goals of separation of control and data plane responsibilities, semantic unification, asynchronous closed-loop, and boundary constraints. The overall architecture is adaptable to various deployment platforms, including user-mode components, kernel-mode processing paths, programmable network interfaces, hardware accelerators, and network processing chips, without being limited to a single deployment form or dependent on a specific operating environment. Specific deployment and operation methods are as follows:

[0039] The semantic layer 110, serving as the core decision input layer, is independently deployed on the network control node and includes a semantic input unit 111, a unified routing decision input unit 113, a routing decision context unit 114, and a routing selection unit 112. During runtime, the semantic input unit 111 continuously retrieves node registration information, session information, and tenant service information from the control bus 120. The unified routing decision input unit 113 structurally integrates various fragmented semantic data, link metrics, and capability summaries to generate routing decision input data in a fixed format. The routing decision context unit 114 caches the decision parameters and status of the current scheduling cycle to ensure scheduling continuity. The routing selection unit 112, combined with static policy interface configuration, completes the final path selection decision. The semantic layer 110 is only responsible for constructing decision inputs and making policy decisions and does not participate in any data plane forwarding operations.

[0040] The control bus 120, serving as the network-wide status management layer, is deployed at the core network control node and includes a registration and heartbeat unit 121, a capability negotiation and policy input unit 122, a control event unit 123, and a static image capture interface 124. The registration and heartbeat unit 121 receives heartbeat messages from all edge nodes and relay nodes every second; if three consecutive heartbeats are not received, the node is considered offline. The capability negotiation and policy input unit 122 completes the interaction and registration of the node's P2P and Relay transmission capabilities when the node comes online. The control event unit 123 generates control events such as node online / offline status, policy changes, and session expiration in real time. The static image capture interface 124 synchronously listens to the routing decision output of the semantic layer 110. This interface relies on a lock-free circular buffer or static shared memory area to capture decision results in a read-only, non-blocking manner for asynchronous archiving and audit analysis, without interfering with the normal scheduling process.

[0041] The data bus 130, serving as the sole data forwarding layer, is distributed across all edge and relay nodes. It comprises a transmission adaptation unit 131, an encapsulation and decapsulation unit 132, a minimum observation unit 133, and a data plane output unit 134. The transmission adaptation unit 131 is compatible with tunnel-based transmission mechanisms, point-to-point direct communication modes, and relay forwarding modes. The tunnel transmission mechanism preferably uses the TUN / TAP interface. The transmission adaptation unit 131 is compatible with multiple transmission modes, including TUN, P2P, Relay, and overlay network encapsulation, and can automatically adapt to different business scenarios. The encapsulation and decapsulation unit 132 is responsible for protocol encapsulation and parsing of user data packets, and for stripping / adding network scheduling metadata. The minimum observation unit 133 collects lightweight metrics such as RTT latency, packet loss rate, and failure count during the forwarding process in real time and synchronizes them to the semantic layer 110. The data plane output unit 134 pushes data packets to the corresponding forwarding path based on routing decisions. The data bus 130 performs only forwarding and lightweight observation throughout the entire process, without policy calculation, state storage, or auditing logic.

[0042] Asynchronous feedback bypass 140 is an independent event processing channel that does not occupy core resources of the data plane or control plane, and has no risk of data forwarding blockage. It receives forwarding results and link status data output from the data plane in real time, generates standardized feedback events, and asynchronously pushes them to the audit module, policy tuning module, and test analysis module. At the same time, it writes back the status update results to the semantic layer 110 routing context unit to achieve scheduling closed loop. Asynchronous feedback bypass 140 only consumes events and writes back the status; it does not modify or intercept the original user data stream.

[0043] In this specific embodiment, the data bus 130 includes at least one of TUN, P2P, Relay, and overlay network encapsulated data plane bearer module, and the overlay network encapsulated data plane bearer module does not assume the responsibility of control plane state management; lightweight link state observation includes at least RTT latency detection, packet loss rate statistics, and path failure count statistics, providing link indicator support for routing decisions.

[0044] In this specific embodiment, the unified routing decision input output by the semantic layer 110 is structured data, including routing status information, relay candidate information, service semantic fields, node capability summaries, and link health indicators. The routing status information includes the current preferred route type, the current active route type, and the maximum hop limit. The relay candidate information includes the primary relay candidate node and the set of backup relay candidate nodes. The service semantic fields characterize the data transmission service attributes, including network affiliation identifier, subject identity identifier, service type identifier, and scheduling priority. The service semantic fields include virtual network identifier, tenant identifier, namespace, service identifier, and service priority. The node capability summary includes whether P2P priority is supported, whether there are valid relay candidates, and the number of healthy relay nodes. The link health indicators include real-time RTT latency, packet loss rate, number of consecutive P2P failures, number of consecutive relay failures, and tunnel degradation status.

[0045] In this specific implementation, the static policy interface is implemented through compile-time pre-configuration, configuration file-driven, or pre-set policy tables. It only supports static policy matching and service profile mapping, and does not include runtime plugin discovery, online hot loading, or dynamic ABI-compatible binding mechanisms. The service profiles that the static policy interface can map include default services, low-latency services, high-volume transmission services, P2P priority scheduling, Relay priority scheduling, and aggressive link recovery.

[0046] In this specific implementation, the asynchronous feedback event includes at least one of the following: P2P transmission failure, P2P transmission success, Relay transmission success, Relay path rollback, priority path recovery, cooling-off period activation, and path health update event. The asynchronous feedback event is bound to the corresponding routing decision result and carries a timestamp. It is output through the asynchronous bypass channel and can be asynchronously consumed by external modules. It does not block or modify the original data plane forwarding data stream and is used to drive path rollback, cooling-off, recovery adaptive logic, and audit analysis.

[0047] In this specific implementation, the system also includes access constraint rules, which are used to delineate the boundary between basic regression baseline capabilities and long-term post-extension capabilities. Basic regression baseline capabilities include closed-loop scheduling based on relay forwarding mode, priority scheduling of point-to-point direct communication mode, and routing rollback and recovery mechanisms when nodes are abnormal, link detection / cooling / recovery, and overlay network encapsulation metadata processing. Basic regression baseline capabilities include single relay closed-loop scheduling, P2P priority scheduling, relay failure rollback, link detection / cooling / recovery, and overlay network encapsulation metadata processing, constituting a stable and regressible delivery baseline for the system. Long-term post-extension capabilities include dynamic extension operation mechanisms, shared memory bus, complete user-space virtual stack, and global strong consistency synchronization mechanisms, which are iterative extension capabilities rather than basic essential capabilities. Long-term post-extension capabilities can only be accessed through static configuration or independent modules, and must reserve degradation and rollback paths, prohibit access to data plane hot paths, and prohibit bypassing the routing selection unit 112 to customize extension scheduling logic.

[0048] Each layer of the three-layer decoupled scheduling architecture of this invention can be independently deployed in user-space processes, kernel-space modules, programmable network interfaces, hardware accelerators, or dedicated network chips. The implementation carrier of the routing decision logic is not limited to a specific computing environment.

[0049] A data scheduling method includes the following steps:

[0050] S1. Control plane status acquisition: Continuously acquires status metadata and routing basic metadata of all network nodes through the control bus 120;

[0051] S2. Unified semantic processing: The semantic layer 110 extracts the business semantic fields from the metadata output by the control bus 120, combines the node capability summary and link health indicators, and generates a standardized unified routing decision input 122.

[0052] S3. Static Policy Matching: Matches the unified routing decision input through the static policy interface to generate a definite target forwarding path;

[0053] S4. Data plane forwarding execution: Data bus 130 completes packet encapsulation, decapsulation and link forwarding according to the target forwarding path, synchronously updates the lightweight link status observation indicators, and triggers the generation of asynchronous feedback events;

[0054] S5. Asynchronous Feedback Event Output: After binding a timestamp to the generated standardized asynchronous feedback event, it is output to the external module for consumption through a non-blocking bypass channel;

[0055] S6. Closed-loop adaptive optimization: External modules consume asynchronous feedback events, update path health status, and drive adaptive iteration of routing scheduling strategies.

[0056] A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks.

[0057] A computer program product, including computer program code, enables the terminal to implement the aforementioned data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks when the computer program product is run on a terminal.

[0058] An electronic terminal includes a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it implements the data scheduling method of the three-layer decoupled scheduling system for distributed overlay networks described above.

[0059] In specific implementations of this invention:

[0060] 1. Specific implementation of static strategy interface:

[0061] The static strategy interface of this invention adopts a dual implementation method of compile-time pre-setting + configuration file solidification, without runtime dynamic loading, plugin registration, or dynamic interface adaptation. The specific implementation process is as follows:

[0062] During the system compilation phase, developers pre-define six basic service profile strategy enumerations, corresponding to default business, low-latency business, high-volume transmission, P2P priority, Relay priority, and aggressive recovery scenarios, respectively. The scheduling rules, path priorities, and threshold parameters for each scenario are solidified into the static code of the program, forming a basic strategy library that cannot be dynamically modified.

[0063] During the system deployment phase, the basic policies can be fine-tuned and adapted through local configuration files. Differentiated policy priorities can be configured for different tenants and different namespaces. For example, real-time audio and video tenants are matched with low-latency service policies by default, and the P2P path with the lowest RTT latency is selected first. File transfer tenants are matched with high-volume transmission policies, and the Relay path with sufficient bandwidth and low packet loss rate is selected first.

[0064] During system operation, the static policy interface only receives unified routing decision input from the semantic layer 110. It performs a one-time policy matching based on the preset policy table and configuration parameters, directly outputting the routing preference result. No dynamic resource loading, interface compatibility checks, or plugin initialization operations are performed throughout the process, completely avoiding the risk of hot path failures caused by dynamic expansion. Furthermore, all policy adjustments require a restart via configuration file or compile-time modifications, with no runtime changes, ensuring the stability and traceability of the scheduling policy.

[0065] 2. Example of constructing unified routing decision input:

[0066] This invention achieves full-dimensional data integration through semantic layer 110, constructing standardized and structured routing decision inputs. The specific construction process is as follows:

[0067] The first step is semantic field collection. When a client node goes online and registers, it actively reports its own virtual network identifier, tenant ID, namespace, service ID, and business priority parameters. The control bus 120 writes the above parameters into the node metadata and session metadata, completing the initial input of business semantics.

[0068] The second step is capability summary collection. The control bus 120 uses the node capability negotiation mechanism to detect whether the current node supports P2P transmission, the number of available relay nodes in the vicinity, and the number of healthy relay nodes, and generates a node capability summary, which is then synchronized to the semantic layer 110.

[0069] The third step is link metric collection. The data bus 130 and the smallest observation unit 133 collect real-time statistics on the real-time RTT latency, packet loss percentage, number of consecutive P2P transmission failures, number of consecutive Relay transmission failures, and tunnel degradation status of each forwarding path, forming a real-time link health metric dataset.

[0070] The fourth step is structured integration and output. The semantic layer 110 integrates the aforementioned semantic fields, capability summaries, and link indicators with the current system configuration's priority route type, active route type, and maximum hop limit parameters to generate a unified route decision input with a fixed structure. This provides a unified and standardized decision-making basis for route selection, completely resolving the problems of semantic fragmentation and chaotic decision-making basis in traditional solutions. The semantic layer 110 of this invention is implemented using a minimal semantic input model, enabling scheduling decisions without relying on a complete business extension system or dynamic semantic operation mechanism.

[0071] 3. Specific implementation of asynchronous feedback closed loop:

[0072] This invention achieves adaptive closed-loop optimization of routing scheduling through standardized feedback events and asynchronous bypass mechanisms. The specific implementation process takes the dynamic switching scenario between P2P and Relay paths as an example:

[0073] Under normal scheduling conditions, the system uses the P2P priority scheduling mode by default, based on the static policy interface configuration. The data bus 130 performs P2P path forwarding, and the smallest observation unit 133 monitors the RTT latency and packet loss rate of the P2P path in real time. When the number of consecutive P2P path failures reaches a preset threshold (3 times), or the real-time packet loss rate exceeds the 5% threshold, the routing selection unit 112 of the semantic layer 110 triggers the path fallback mechanism, switching the data forwarding path from the P2P path to the backup relay path.

[0074] After the path switch is completed, the data plane proactively generates two types of standardized feedback events: "P2P transmission failed" and "Relay fallback effective." These events are bound to the current routing decision number, a high-precision timestamp, and real-time link metrics data, and are pushed to the backend auditing and policy tuning modules without loss via asynchronous feedback bypass 140. Simultaneously, the system automatically triggers a P2P path cooling mechanism. Within a preset dynamic time window \Delta T (preferably 5s - 30s), the system is prohibited from actively switching back to the P2P priority path during the cooling period to avoid frequent path oscillations.

[0075] During the cooling-off period, the data bus 130 continuously monitors the health status of the P2P path. If the P2P path RTT latency and packet loss rate are detected to return to the normal threshold range, a "path health recovery" feedback event is automatically generated. After receiving the event, the semantic layer 110 removes the cooling-off restriction and switches the forwarding path from the Relay path back to the P2P priority path, while recording the "priority path recovery" event.

[0076] All feedback events are processed asynchronously, only recording status and updating policies, without participating in data plane forwarding logic. This ensures no forwarding delays, link blockages, or jitter throughout the process. The backend auditing module can perform fault tracing and compliance recording based on event data, the policy module can optimize threshold parameters based on historical event data, and the test analysis module can perform link quality assessment and regression verification based on event sequences, achieving adaptive iterative optimization of scheduling policies and a fully auditable closed loop across the entire link.

[0077] 4. Specific implementation of access boundary constraints:

[0078] This invention strictly defines the architectural capability boundaries through access constraint rules, achieving isolation between basic baseline capabilities and long-term scalability capabilities. The specific implementation method is as follows:

[0079] The system defines the basic regression baseline capabilities and incorporates single relay closed-loop scheduling, P2P priority scheduling, relay failure fallback, link detection / cooling / recovery, and overlay network encapsulation metadata processing as core basic capabilities into the system's default operating logic, supporting normalized stable delivery and automated regression testing.

[0080] As another preferred embodiment of the present invention (hardware acceleration form), the system further includes a dynamic expansion mechanism, wherein the control bus and data bus communicate via a shared-memory bus or a user-space virtual protocol stack, thereby further compressing the main path scheduling overhead to the line-speed level of the hardware network interface card (DPU / SmartNIC). The dynamic expansion mechanism, shared-memory bus, complete user-space virtual stack, and global strong consistency synchronization mechanism are defined as long-term post-expansion capabilities, and direct access of such capabilities to the data plane hot path is prohibited. If services require the addition of auditing, policy arbitration, AI parameter tuning, or other expansion capabilities, they must be accessed through independent modules and static configuration switches, and the core three-layer architecture and data plane forwarding logic must not be modified.

[0081] Meanwhile, the system has built-in automated verification rules and code review mechanisms to detect in real time any violations such as backend capability intrusion into hot paths, dynamic plugin loading, or bypassing routing unit 112 to extend custom logic. If a violation is detected, a downgrade interception is automatically triggered to ensure the stability and standardization of the architecture boundaries. All newly added extended capabilities declare independent stage identifiers, function switches, regression test matrices, and a one-click rollback mechanism, which can quickly restore the basic baseline capability state and avoid the delivery and review risks caused by long-term capability pre-positioning.

[0082] 5. Static Image Capture Interface Example:

[0083] The static image capture interface 124 integrated into the control bus 120 of this invention adopts a read-only copy mechanism. After the routing unit 112 completes its decision, it synchronously writes the routing decision summary and service semantic information into the memory-mapped area (or circular buffer). The control bus 120 achieves non-blocking capture by subscribing to data in this area, without occupying data plane forwarding thread resources throughout the process. The specific implementation process is as follows:

[0084] After the semantic layer 110 routing selection unit 112 completes the routing decision and generates the final forwarding path parameters, it synchronously calls the static image capture interface 124 built into the control bus 120. This interface only reads and copies the routing decision results, service semantic information, and link state summary data, without modifying the original decision parameters and user data packet content, and does not occupy data plane forwarding thread resources, and runs non-blockingly throughout the process.

[0085] The interface pushes captured structured data to an asynchronous audit queue for offline debugging, scheduling effect analysis, compliance auditing, and fault tracing, providing a standardized data source for bypass auditing, problem review, and strategy iteration. This interface can be implemented using equivalent structures such as compile-time static function tables, fixed interface structures, and configuration tables, requiring no runtime adaptation and adapting to all deployment environments, including user-space, kernel-space, and hardware-accelerated deployments. All performance parameters, threshold ranges, and scenario descriptions in this invention are for illustrative purposes only and do not constitute production SLAs, deterministic QoS, or commercial delivery commitments.

[0086] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. If those skilled in the art are inspired by this description and design similar invention content and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A three-layer decoupled scheduling system for distributed overlay networks, characterized in that, The system comprises a data bus, a control bus, and a semantic layer. The data bus is responsible for packet carrying, encapsulation, decapsulation, forwarding, and lightweight link state observation. This lightweight link state observation includes, but is not limited to, one or more indicators based on latency, bit error rate or packet loss rate, and connection stability. It only handles data plane forwarding and does not participate in control plane state management or policy judgment. The control bus is responsible for node registration, heartbeat detection, node discovery, session maintenance, route hints, capability negotiation, and control event output. It also integrates a static image capture interface, which uses a preset lock-free circular buffer or static shared memory area to capture and output routing decision results and associated business semantic information in a non-blocking manner. The semantic layer is an independent semantic processing layer that uses a minimal semantic input model to convert tenant identifiers, namespaces, service identifiers, business priorities, and node capability summaries into unified routing decision inputs. The system completes business policy matching through a static policy interface and achieves adaptive closed-loop optimization of routing policies through asynchronous feedback events.

2. The three-layer decoupled scheduling system for distributed overlay networks according to claim 1, characterized in that, The data bus includes at least one of an overlay network transmission adapter and an overlay network encapsulated data plane bearer module, and the data plane bearer module does not assume the responsibility of control plane state management; the lightweight link state observation includes at least RTT latency detection, packet loss rate statistics, and path failure count statistics, providing link indicator support for routing decisions.

3. A three-layer decoupled scheduling system for distributed overlay networks according to claim 2, characterized in that, The adaptation methods supported by the overlay network transmission adapter include one or more of the following: tunnel-based transmission mechanism, point-to-point (P2P) direct communication mode, or relay forwarding mode; wherein, the tunnel-based transmission mechanism is preferably the TUN / TAP interface.

4. A three-layer decoupled scheduling system for distributed overlay networks according to claim 3, characterized in that, The unified routing decision input output by the semantic layer is structured data, including routing status information, relay candidate information, service semantic fields, node capability summaries, and link health indicators. The routing status information includes the current preferred route type, the current active route type, and the maximum hop limit. The relay candidate information includes the primary relay candidate node and the set of backup relay candidate nodes. The service semantic fields are used to characterize the service attributes of data transmission, including but not limited to virtual network identifiers, tenant identifiers, namespaces, service identifiers, and service priorities. The node capability summaries include whether P2P priority is supported, whether there are valid relay candidates, and the number of healthy relay nodes. The link health indicators include real-time transmission latency, link loss rate, and tunnel degradation status.

5. A three-layer decoupled scheduling system for distributed overlay networks according to claim 3, characterized in that, The static policy interface is implemented through compile-time pre-configuration, configuration file-driven, or pre-set policy tables. It only supports static policy matching and business profile mapping, and does not include runtime plugin discovery, online hot loading, or dynamic ABI-compatible binding mechanisms. The service profiles that the static policy interface can map include default services, low-latency services, high-volume transmission services, P2P priority scheduling, Relay priority scheduling, and aggressive link recovery.

6. A three-layer decoupled scheduling system for distributed overlay networks according to claim 3, characterized in that, The asynchronous feedback event includes at least one of the following: P2P transmission failure, P2P transmission success, Relay transmission success, Relay path rollback, priority path recovery, cooling-off period activation, and path health update. The asynchronous feedback event is bound to the corresponding routing decision result and carries a timestamp. It is output through an asynchronous bypass channel and can be asynchronously consumed by external modules. It does not block or modify the original data plane forwarding data stream and is used to drive path rollback, cooling-off, recovery adaptive logic, and audit analysis.

7. A three-layer decoupled scheduling system for distributed overlay networks according to claim 6, characterized in that, The external module consumes the asynchronous feedback event in a non-blocking manner through an asynchronous bypass channel or a local lightweight interface, and the consumption process does not modify or intercept the original data plane forwarding stream.

8. A three-layer decoupled scheduling system for distributed overlay networks according to claim 3, characterized in that, The system also includes access constraint rules, which define the boundary between basic regression baseline capabilities and long-term post-extension capabilities. These rules are configured with a routing unit interception strategy. When a data stream containing dynamic extension registration instructions or runtime plugin loading requests is detected, the routing unit intercepts the request and directs it to an independent module for execution, preventing resource preemption of the data plane forwarding thread. The basic regression baseline capabilities include closed-loop scheduling based on relay forwarding mode, priority scheduling for point-to-point direct communication mode, route rollback and recovery mechanisms in case of node anomalies, link detection / cooling / recovery, and overlay network encapsulation metadata processing, forming a stable and regressible delivery baseline. The long-term post-extension capabilities include a dynamic extension operation mechanism, shared memory bus, complete user-space virtual stack, and a global strong consistency synchronization mechanism; these are iterative extension capabilities rather than essential basic capabilities. Long-term post-extension capabilities can only be accessed through static configuration or independent modules, and must reserve degradation and rollback paths. Access to data plane hot paths is prohibited, and custom extension scheduling logic cannot be bypassed by the routing unit.

9. A data scheduling method, characterized in that, Applied to a three-layer decoupled scheduling system for distributed overlay networks as described in any one of claims 1-8, the method includes the following steps: S1. Control plane status acquisition: Continuously acquire status metadata and routing basic metadata of all network nodes through the control bus; S2. Unified semantic processing: The semantic layer extracts the business semantic fields from the metadata output by the control bus, combines the node capability summary and link health indicators, and generates standardized unified routing decision input; S3. Static policy matching: The unified routing decision input is matched through the static policy interface to generate a determined target forwarding path; S4. Data plane forwarding execution: The data bus completes packet encapsulation, decapsulation and link forwarding according to the target forwarding path, synchronously updates the lightweight link status observation indicators, and triggers the generation of asynchronous feedback events; S5. Asynchronous Feedback Event Output: After binding a timestamp to the generated standardized asynchronous feedback event, it is output to the external module for consumption through a non-blocking bypass channel; S6. Closed-loop adaptive optimization: The external module consumes the asynchronous feedback events, updates the path health status, and drives the adaptive iteration of the routing scheduling strategy.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks as described in claim 9.

11. A computer program product, comprising computer program code, characterized in that, When the computer program product is run on a terminal, the terminal implements the data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks as described in claim 9.

12. An electronic terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the data scheduling method for a three-layer decoupled scheduling system for distributed overlay networks as described in claim 9.

Images