A multi-modal network dynamic loading and management system and method based on modal virtualization

Abstract

This invention relates to a multimodal network dynamic loading and management system and method based on modal virtualization, belonging to the field of network architecture and virtualization technology. By constructing a modal virtualization layer and a modal description language parsing mechanism, it breaks through the limitation of traditional network virtualization that only faces a single protocol stack, realizing unified encapsulation and isolation of multiple network modes such as IP, TSN, ICN, and custom protocols. By utilizing dual-modal state snapshot and traffic mirroring technology, it solves the problem of service interruption during online switching of network modes. Furthermore, by combining a vectorized heterogeneous resource elastic scheduling model, it realizes global optimized configuration and on-demand mapping of CPU, GPU, FPGA, and P4 programmable switching chip.

Description

Technical Field

[0001] This invention relates to a multimodal network dynamic loading and management system and method based on modal virtualization, belonging to the field of network architecture and virtualization technology. Background Technology

[0002] With the rapid development of information and communication technologies, network architectures are gradually shifting from single-protocol and fixed-structure architectures to diversified, programmable, and evolvable architectures. The parallel development of 5G, WiFi 6, TSN (Time-Sensitive Networking), ICN (Information-Centric Networking), and various dedicated communication protocols has led to a trend of multimodal, multi-access, and multi-protocol coexistence in networks. Simultaneously, the deep integration of cloud computing and edge computing is driving the coordinated evolution of the "cloud-network-edge-device" integration, requiring network infrastructure to possess greater flexibility and reconfigurability. However, existing network architectures are mostly single-modal and fixed designs, or rely on Network Functions Virtualization (NFV) for traditional resource layer abstraction (such as virtual machines and containers). Existing technologies have the following significant drawbacks when facing multimodal environments with public cloud network resources:

[0003] Virtualization granularity mismatch: Existing technologies focus on the segmentation of computing or link resources, making it difficult to achieve overall encapsulation based on "network modalities". Due to the huge differences in protocol stack structure, message format and forwarding logic between different modalities (such as IP and TSN), traditional methods cannot achieve unified encapsulation and isolation of multimodal systems.

[0004] Difficulty in coordinating heterogeneous resources: In the public cloud network environment, network elements contain a variety of heterogeneous resources such as CPU, GPU, FPGA and programmable forwarding chip. Existing solutions are difficult to establish a unified abstract model and cannot realize on-demand mapping and elastic reuse of heterogeneous resources among multiple modes.

[0005] Lack of dynamic evolution capability: Existing orchestration mechanisms mostly rely on static configuration and full replacement, lacking the ability to dynamically load, seamlessly switch and incrementally evolve modes at runtime, making it difficult to support smooth upgrades and verification under real network conditions.

[0006] Therefore, there is an urgent need for an architecture that can virtualize, encapsulate, dynamically load, and manage the entire lifecycle of multiple network modes on a unified physical environment and heterogeneous resources, in order to meet the complex requirements of future 6G networks, multi-access edge computing, and cloud-network converged environments. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a dynamic loading and management system and method for multimodal networks based on modal virtualization. By constructing a modal virtualization layer (MVL) and a lossless dual-modal switching mechanism, it overcomes the limitations of traditional virtualization technology, which is only geared towards computing or link resource partitioning, and realizes dynamic loading, co-operation, and evolution management of multimodal networks in a public cloud network environment.

[0008] Terminology Explanation:

[0009] Public cloud network resources refer to the underlying physical infrastructure that constitutes the system. This includes not only virtualized computing resources, but also physical network devices (such as physical servers, physical switches that support P4 programming, edge gateways, etc.) and heterogeneous acceleration hardware (FPGA cards, GPU cards).

[0010] Modality: refers to a virtualized instance of a network protocol stack or communication mode that can run independently (such as IP, TSN, or ICN).

[0011] The technical solution of the present invention is as follows:

[0012] A multimodal network dynamic loading and management system based on modal virtualization includes:

[0013] The physical resource layer, namely the heterogeneous public cloud network resource pool, serves as the system's bottom layer and includes: general-purpose computing resources (CPU cores), accelerated computing resources (GPU cores, used for matrix operations and deep learning inference), programmable logic units (FPGAs, used for hardware acceleration), and programmable forwarding chips (used for customizing data plane forwarding logic, such as P4 switches that support protocol-independent programming, which allow for defining packet parsing and forwarding behavior for non-standard network protocols through programming). Each resource is configured with real-time monitoring capabilities, and utilization and load status are fed back in real time through a unified abstract interface for heterogeneous resources.

[0014] Modal Virtualization Layer (MVL) is used to establish a unified abstract model for heterogeneous resources, execute resource scheduling algorithms, manage the lifecycle of modalities, and maintain a global dependency view.

[0015] At the application layer, a multimodal network runtime environment is established. Each modality is encapsulated in an independent virtualized container, with runtime support provided by a modal container runtime engine. Container technology can be flexibly selected based on the security and performance requirements of each modality. For example, Docker containers are used for high-performance modalities, while Kata Containers lightweight virtual machines are used for highly isolated modalities. The specific configuration of the container (such as CPU core binding and memory quota) is directly derived from the resource requirements definition in the modal MDL requirement file and is automatically injected by the system at runtime.

[0016] Modal dynamic loading and switching controller: Located above the modal virtualization layer, it is used to issue scheduling instructions to the modal virtualization layer. It consists of a scheduler, a status monitor and an event processor. The three are logically connected through an internal message bus to form a closed-loop control system of perception-decision-execution.

[0017] According to a preferred embodiment of the present invention, in the modal virtualization layer:

[0018] a. Establish a unified abstract model for heterogeneous resources. The modal virtualization layer has a built-in resource adapter. Figure 1 The "Heterogeneous Resource Unified Abstraction Interface" shown reads the physical parameters of the physical resource layer (such as CPU clock speed, GPU memory, number of FPGA logic units, P4 flow table capacity, etc.), and uses a standardized mapping mechanism to normalize the physical parameters into a unified virtual resource descriptor, thereby shielding the differences in the underlying hardware and building a globally unified virtual resource pool view.

[0019] b. Execute resource scheduling algorithm: The modal virtualization layer integrates a scheduling solver interface, collects the remaining capacity status of the physical resource layer in real time, and combines it with the MDL requirement file of the modality to be loaded to transform the scheduling problem into a matrix mathematical model. The modal virtualization layer calls the built-in scheduling solver (such as an algorithm engine based on branch and bound) to calculate the optimal resource allocation decision matrix (X) and generate deployment instructions for specific nodes of the physical resource layer accordingly.

[0020] c. Managing the lifecycle of modes: The modal virtualization layer maintains a finite state machine (FSM) to manage the state transitions of each mode, including:

[0021] Creation phase: Parse the MDL requirement file of the modality, pull the corresponding image or bitstream file, call the underlying container engine to start the modality, and issue hardware configuration instructions (such as loading FPGA logic or P4 flow table).

[0022] Pause / Migration Phase: Triggers the state snapshot mechanism, saves the context, and freezes the container;

[0023] Destruction phase: Reclaim computing resources, clean up flow table entries, and update the availability status of the global resource pool;

[0024] d. Maintain a global dependency view: Based on the dependency graph (DAG) generated by parsing the modal MDL requirement file, the modal virtualization layer monitors the health status of each dependency (such as clock service, underlying link) in real time to ensure that the prerequisites for modal operation are met.

[0025] According to a further preferred embodiment of the present invention, the state snapshot mechanism is as follows: when a mode is paused or migrated, a snapshot (running state) of the old mode needs to be saved, and the new mode is initialized with this snapshot. The state transition follows the following formula:

[0026]

[0027] in, S represents the initial running state when the new mode is started; snapshot This represents the set of state snapshots captured at the switching time t0 of the old mode, specifically including the flow table (F flow ), buffer queue (Q) queue Session control context (C) ctrl (such as TCP connection tables, sequence number status, etc.) and metadata (M meta ); S config This formula represents the basic static configuration of the new mode (such as routing tables and QoS policy templates). It indicates that the new mode inherits the complete business context of the old mode, ensuring that existing session connections will not be disconnected due to reset after the switch.

[0028] According to a preferred embodiment of the present invention, in the modal dynamic loading and switching controller:

[0029] The status monitor is used to perceive and maintain the global resource status in real time. It periodically polls the utilization rate and operation indicators of each mode (such as latency and packet loss rate) of the heterogeneous public cloud network resource pool through the unified abstract interface of heterogeneous resources, and updates this real-time data to the shared memory for the scheduler to query.

[0030] Event handlers, used for coordination: As the signaling entry point of the system, they are responsible for receiving external modal loading requests (MDL requirement files) or internal abnormal alarms triggered by the monitor (such as heartbeat loss). Once an event is received, the event type is parsed and a scheduling request is sent to the scheduler.

[0031] The scheduler, used for decision-making, acts as the core decision engine. Upon receiving a request from an event handler, the scheduler immediately reads the current global resource status from the status monitor, combines it with the requirement constraints in the modal MDL requirement file, and executes a matrix scheduling algorithm or a lossless switching strategy to generate the optimal resource allocation scheme (decision matrix). The command, or switching instruction, is ultimately sent to the modal virtualization layer for execution.

[0032] A further preferred embodiment of the lossless switching strategy according to the present invention is:

[0033] The modal dynamic loading and switching controller first parses the MDL requirement file and generates a dependency graph. After passing resource constraint checks (if not met, a queuing or degradation strategy is triggered), a new modal container is created and the state snapshot of the old modality is restored. Then, it enters the dual-modal parallel stage until the state synchronization is completed, at which point atomic switching is performed. The specific strategy and algorithm logic are as follows:

[0034] S1. Dual-modal state merging: During lossless switching, the system allows the old and new modes to coexist for a short period of time (dual-modal parallelism), within the switching window. Within the modality, the dynamic loading and switching controller issues commands to execute traffic mirroring, simultaneously distributing inbound traffic to both the new and old modalities. During this period, the new modality not only has an initial snapshot state but also processes the mirrored traffic in real time to generate new state increments. The state merging logic is as follows:

[0035]

[0036] This mechanism ensures that incremental data generated during handover (such as newly arrived packets) is not lost. In the formula, This represents the complete state of the new mode at any time t within the switching window; The initial state is calculated using the aforementioned formula; During parallel operation, the state increment generated by the mirrored traffic in the new mode is processed in real time. The dual-modal parallel phase continues until the state difference between the old and new modes converges to a preset threshold (such as state synchronization rate > 99%), and then an atomic switching operation is performed to flip the traffic pointer.

[0037] S2. Switching Window Optimization Constraints: To ensure service quality, the switching process must meet the following optimization objectives:

[0038]

[0039] in, P represents the duration of the window switching. loss This represents the packet loss rate during the handover period. This represents the maximum allowable packet loss rate threshold of the system (e.g.: This constraint ensures that while pursuing rapid switching, the system must prioritize meeting the reliability requirements of the business.

[0040] According to a preferred embodiment of the present invention, a heterogeneous resource elastic scheduling and QoS guarantee mechanism is configured within the modal dynamic loading and switching controller:

[0041] ① Basic resource allocation model: The scheduler aims to maximize the total system utility, and the following optimization model is established:

[0042]

[0043] Among them, u i,k x represents the utility weight of mode i using resource k. i,k This is a Boolean variable, with a value of 1 representing allocation and 0 representing no allocation. It must satisfy the total capacity constraint. d i,k C represents the unit requirement of resource k for running mode i. k Let k be the total physical capacity of resource k.

[0044] This weight supports dynamic updates: the system dynamically adjusts the u of each modality based on real-time monitoring of business SLA achievement rates. i,k The specific adjustment logic is as follows: when monitoring data shows that the SLA indicator (such as latency or packet loss rate) of a certain modality i is close to or worse than a preset threshold, the system automatically increases its corresponding utility weight u. i,k (e.g., execute) , (This is used as a priority boosting factor), thereby increasing the priority of the resources allocated to this mode in the next round of optimization, forcing the scheduler to prioritize the resource needs of high-value or sub-optimal businesses, and triggering online rescheduling;

[0045] ② Vectorized / Matrix Representation of the Scheduling Problem: In actual engineering deployments, in order to use linear programming solvers (such as Gurobi or CPLEX) to quickly solve large-scale nodes, the scheduling problem is modeled in a vectorized and matrix manner. The optimization model belongs to a typical mixed integer linear programming (MILP) problem, which is solved efficiently by commercial solvers. It is assumed that the number of modes to be scheduled in the system is N, and the number of heterogeneous resource types involved is K.

[0046] (21) Define the resource allocation decision matrix :

[0047]

[0048] Among them, matrix elements Let them be Boolean decision variables. , This indicates that the i-th mode has been allocated the k-th type of resource, and 0 indicates that it has not been allocated;

[0049] (22) Define the resource demand moment D and the capacity vector C: Let This is a resource demand matrix, where, Let represent the unit demand of the i-th mode for the k-th type of resource; let This is a resource capacity vector, a column vector representing the total capacity of each of the K types of resources in the system, indicating the upper limit of the physical total amount of each resource.

[0050] (23) Matrix-based constraints: Global resource constraints are uniformly represented as:

[0051]

[0052] in, This represents the Hadamard product of matrices (i.e., element-wise multiplication), a form that directly preserves the modality-resource correspondence and avoids cross-resource interference. This yields an N×K intermediate matrix, where non-zero elements represent the actual amount of resources occupied, and T represents the matrix transpose operation, which transposes the N×K intermediate matrix into a K×N matrix, arranging the occupancy status of the same type of resources in the same row. It is a column vector of dimension N consisting entirely of 1s;

[0053] The engineering physical meaning of this constraint formula is: the system monitors in real time the cumulative occupancy of all currently running modes on various heterogeneous resources such as CPU, GPU, and FPGA, ensuring that it is strictly limited within the total hardware capacity of the physical nodes, thereby fundamentally avoiding system overload or crash due to resource overselling.

[0054] (24) Matrix-based objective function: total system utility U total Represented as:

[0055]

[0056] in, (i.e., on the real number field) The dimensional matrix is ​​a predefined utility weight matrix. It is a dimension of A row vector of all 1s; For dimension A column vector of all 1s; This represents the utility value matrix of the selected allocation schemes. ;

[0057] This formula uses matrix operations to calculate the total utility value of all modes and resource types under the current scheduling scheme in one go, which makes it easy for computers to solve efficiently using the branch and bound method or existing integer programming algorithms.

[0058] According to a preferred embodiment of the present invention, the system design modality definition and parsing mechanism, in order to achieve unified definition and parsable loading of modalities, the system designs a modality description language, which is implemented by the MDL parser built into the Modality Virtualization Layer (MVL), serving as the standard input interface for MVL to identify and manage network modalities. The MDL requirement file is usually in JSON or YAML format, covering fields such as protocol stack composition, message structure, routing policy, QoS constraints, resource requirements, and security policy;

[0059] This includes defining examples and syntactic constraints;

[0060] Parsing process and dependency graph construction: The modal virtualization layer calls the MDL parser to read the file and generate modal metadata objects in memory. The MDL parser automatically analyzes the Dependencies field and generates a dependency graph—represented by a Directed Acyclic Graph (DAG) structure. It is then merged into the global dependency view. The specific dependency types include: hardware dependencies (such as the need to exclusively use a specific FPGA card), service dependencies (such as the need to rely on the underlying clock synchronization service), and logical dependencies (such as the ICN modality depending on the IPUnderlay modality). If a dependency is missing or physical resources are insufficient, the modal dynamic loading and switching controller will trigger a queuing or resource preemption strategy.

[0061] According to a preferred embodiment of the present invention, in order to overcome the performance bottleneck of traditional virtualization, the system constructs a deep software-hardware co-mapping mechanism:

[0062] First, the underlying physical network interface card (NIC) uses SR-IOV technology to virtualize multiple virtual functions (VFs) and directly mounts them as virtual interfaces (vNICs) inside the container through a hardware mapping mechanism, achieving zero-copy access to the data plane. Second, the underlying FPGA acceleration card (such as Slot 0) uses PCIe passthrough technology to bypass the host operating system and directly interface with the packet parser inside the container, enabling it to directly call hardware logic for line-speed packet classification and unpacking, significantly reducing processing latency. At the same time, the system uses CPU affinity binding technology to lock the underlying dedicated CPU cores (such as Core 3-4) to serve the protocol stack inside the container, ensuring that computing tasks exclusively use physical cores and avoiding jitter caused by context switching. Thus, an efficient pipelined operation structure is formed inside the container: virtual interface reception - packet parser accelerated processing - protocol stack logic operation - control agent status reporting, achieving near-bare-metal performance while ensuring virtualization isolation.

[0063] According to a preferred embodiment of the present invention, the system incorporates a robust processing strategy:

[0064] I. Severe Resource Shortage: When the loading of a new modality causes the cumulative resource usage of various types to exceed the threshold, i.e., the following judgment formula is satisfied:

[0065]

[0066] Where F represents the total number of modes currently in operation; This represents the resource requirement k for the new modality to be loaded; C k Given the total physical capacity of resource k, when the above inequality holds, the system executes the following sequentially:

[0067] Ⅰ-1. Global Scan: Locate free nodes for migrated containers;

[0068] Ⅰ-2. Resource Degradation: While ensuring the minimum service level agreement (SLA), temporarily reduce redundant resources in low-priority modalities (such as rolling back GPU acceleration allocated to non-critical modalities to CPU computing) to free up resources for high-priority modalities;

[0069] Ⅰ-3. Waiting in a queue: If the above steps are ineffective, the new modality is placed in the waiting queue;

[0070] II. Multimodal Concurrent Switching Protection: When multiple modes trigger switching simultaneously, the system limits the number of concurrent switches to protect control plane performance. ,in, M is the size of the current concurrent switching set. max The maximum number of concurrent switching allowed by the system is set, and transaction management is performed on batch switching: if any mode switching fails, the state of all related modes in the entire transaction is automatically rolled back;

[0071] Ⅲ. Container Anomaly Self-Healing: The self-healing mechanism is automatically triggered after the state monitor detects an abnormal event. Abnormal events include, but are not limited to: unexpected exit of container process, loss of heartbeat, resource exhaustion (OOM), and deadlock detection. After an anomaly is detected, the system automatically rolls back to the most recent stable state snapshot and restarts the container.

[0072] IV. Adaptive Adjustment of Monitoring Thresholds: To further improve the robustness of the system, an adaptive heartbeat detection mechanism is introduced. The system dynamically adjusts the heartbeat detection cycle and abnormal trigger threshold for modal containers based on the current overall load trend and historical operating data. For example, during periods of high network load and congestion, the timeout judgment threshold is appropriately relaxed to avoid false triggering of the self-healing mechanism due to brief network jitter, thereby achieving a balance between sensitivity and stability.

[0073] The four core modules of this invention (physical resource layer, modal virtualization layer, application layer, and modal dynamic loading and switching controller) are tightly coupled with a logical bus through well-defined interfaces, forming a closed-loop operating system. Firstly, at the basic support level, the physical resource layer and the modal virtualization layer (MVL) establish a bidirectional relationship of "abstraction and pooling": various heterogeneous hardware at the bottom layer collects real-time status data through probes and reports upwards through the "heterogeneous resource unified abstraction interface" of the MVL layer. The MVL then uses its built-in resource adapter to normalize the heterogeneous data into virtual resource descriptors, achieving pooling of physical resources. Conversely, the MVL translates upper-layer logical operations (such as container startup) into specific hardware instructions (such as PCIe channel configuration) and sends them to the physical resource layer for execution. Based on this, a "support and bearer" relationship is formed between the MVL and the multimodal network operating environment: the "modal container runtime engine" within the MVL directly supports the application layer's operating environment, responsible for creating container instances, mapping or directly connecting the underlying physical resources to the container, and maintaining the isolation and state consistency of this environment. Secondly, at the management and decision-making level, the Modal Dynamic Loading and Switching Controller (MDL) and MVL constitute a closed-loop interactive system of "decision and execution." The MDL acts as the "commander" above the MVL, while the MVL acts as the "executor," responsible for specific implementation. This closed-loop process begins with perception, where the MVL synchronizes the global resource view and modal operating status to the controller's "status monitor" in real time; then, decision-making is made, with the controller combining MDL requests and resource status scheduling algorithms to generate a decision matrix (…). The command is executed and sent to the MVL, which parses and calls the underlying interface to complete the actual resource allocation or traffic switching operation, thereby achieving dynamic and precise control over the entire multimodal network system.

[0074] A method for dynamic loading and management of multimodal networks based on modal virtualization, comprising the following steps:

[0075] (1) Utilize the heterogeneous public cloud network resource pool to provide feedback on utilization rate and load status;

[0076] (2) Establish a unified abstract model for heterogeneous resources, adopt a standardized mapping mechanism, normalize the physical parameters of heterogeneous public cloud network resource pools into a unified virtual resource descriptor, thereby shielding the differences in underlying hardware, constructing a globally unified virtual resource pool view, then executing resource scheduling algorithms, managing the lifecycle of modalities, and maintaining a global dependency view.

[0077] (3) Maintain global resource status in real time, poll the resource utilization rate and operation indicators (such as latency and packet loss rate) of the heterogeneous public cloud network resource pool, and update these real-time data to the shared memory;

[0078] (4) Receive external events, modal loading requests or internal abnormal alarms triggered by the monitor (such as heartbeat loss). After receiving the event, read the current global resource status, execute the matrix scheduling algorithm or lossless switching strategy, generate the optimal resource allocation scheme or switching instruction, and then issue it for execution.

[0079] The beneficial effects of this invention are as follows:

[0080] This invention overcomes the limitations of traditional network virtualization, which only targets a single protocol stack, by constructing a Modal Virtualization Layer (MVL) and a Modal Description Language (MDL) parsing mechanism. It achieves unified encapsulation and isolation of multiple network modes, such as IP, TSN, ICN, and custom protocols. By utilizing dual-modal state snapshot and traffic mirroring technologies, it solves the problem of service interruption during online network mode switching. Furthermore, by combining a vectorized heterogeneous resource elastic scheduling model, it realizes global optimized configuration and on-demand mapping of CPU, GPU, FPGA, and P4 programmable switching chips. This system has good scalability and real-world verifiability, providing a reference implementation framework for the multimodal evolution of 6G and cloud-network converged networks. Attached Figure Description

[0081] Figure 1 This is a system architecture diagram of the present invention;

[0082] Figure 2 This is a microscopic internal structure diagram of a single modal container of the present invention;

[0083] Figure 3 This is a flowchart of the modal dynamic loading and dual-modal lossless switching of the present invention;

[0084] Figure 4 This is a schematic diagram illustrating the matrix modeling and mapping principle of the heterogeneous resource elastic scheduling of the present invention. Detailed Implementation

[0085] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0086] Example 1:

[0087] like Figure 1 As shown, this embodiment provides a multimodal network dynamic loading and management system based on modal virtualization, including:

[0088] The physical resource layer, namely the heterogeneous public cloud network resource pool, serves as the system's bottom layer and includes: general-purpose computing resources (CPU cores), accelerated computing resources (GPU cores, used for matrix operations and deep learning inference), programmable logic units (FPGAs, used for hardware acceleration), and programmable forwarding chips (used for customizing data plane forwarding logic, such as P4 switches that support protocol-independent programming, which allow for defining packet parsing and forwarding behavior for non-standard network protocols through programming). Each resource is configured with real-time monitoring capabilities, and utilization and load status are fed back in real time through a unified abstract interface for heterogeneous resources.

[0089] Modal Virtualization Layer (MVL) is used to establish a unified abstract model for heterogeneous resources, execute resource scheduling algorithms, manage the lifecycle of modalities, and maintain a global dependency view.

[0090] At the application layer, a multimodal network runtime environment is established. Each modality is encapsulated in an independent virtualized container, with runtime support provided by a modal container runtime engine. Container technology can be flexibly selected based on the security and performance requirements of each modality. For example, Docker containers are used for high-performance modalities, while Kata Containers lightweight virtual machines are used for highly isolated modalities. The specific configuration of the container (such as CPU core binding and memory quota) is directly derived from the resource requirements definition in the modal MDL requirement file and is automatically injected by the system at runtime.

[0091] Modal dynamic loading and switching controller: Located above the modal virtualization layer, it is used to issue scheduling instructions to the modal virtualization layer. It consists of a scheduler, a status monitor and an event processor. The three are logically connected through an internal message bus to form a closed-loop control system of perception-decision-execution.

[0092] In the modal virtualization layer:

[0093] a. Establish a unified abstract model for heterogeneous resources. The modal virtualization layer has a built-in resource adapter. Figure 1 The "Heterogeneous Resource Unified Abstraction Interface" shown reads the physical parameters of the physical resource layer (such as CPU clock speed, GPU memory, number of FPGA logic units, P4 flow table capacity, etc.), and uses a standardized mapping mechanism to normalize the physical parameters into a unified virtual resource descriptor, thereby shielding the differences in the underlying hardware and building a globally unified virtual resource pool view.

[0094] b. Execute resource scheduling algorithm: The modal virtualization layer integrates a scheduling solver interface, collects the remaining capacity status of the physical resource layer in real time, and combines it with the MDL requirement file of the modality to be loaded to transform the scheduling problem into a matrix mathematical model. The modal virtualization layer calls the built-in scheduling solver (such as an algorithm engine based on branch and bound) to calculate the optimal resource allocation decision matrix (X) and generate deployment instructions for specific nodes of the physical resource layer accordingly.

[0095] c. Managing the lifecycle of modes: The modal virtualization layer maintains a finite state machine (FSM) to manage the state transitions of each mode, including:

[0096] Creation phase: Parse the MDL requirement file of the modality, pull the corresponding image or bitstream file, call the underlying container engine to start the modality, and issue hardware configuration instructions (such as loading FPGA logic or P4 flow table).

[0097] Pause / Migration Phase: Triggers the state snapshot mechanism, saves the context, and freezes the container;

[0098] Destruction phase: Reclaim computing resources, clean up flow table entries, and update the availability status of the global resource pool;

[0099] d. Maintain a global dependency view: Based on the dependency graph (DAG) generated by parsing the modal MDL requirement file, the modal virtualization layer monitors the health status of each dependency (such as clock service, underlying link) in real time to ensure that the prerequisites for modal operation are met.

[0100] According to a further preferred embodiment of the present invention, the state snapshot mechanism is as follows: when a mode is paused or migrated, a snapshot (running state) of the old mode needs to be saved, and the new mode is initialized with this snapshot. The state transition follows the following formula:

[0101]

[0102] in, S represents the initial running state when the new mode is started; snapshot This represents the set of state snapshots captured at the switching time t0 of the old mode, specifically including the flow table (F flow ), buffer queue (Q) queue Session control context (C) ctrl (such as TCP connection tables, sequence number status, etc.) and metadata (M meta ); S config This formula represents the basic static configuration of the new mode (such as routing tables and QoS policy templates). It indicates that the new mode inherits the complete business context of the old mode, ensuring that existing session connections will not be disconnected due to reset after the switch.

[0103] In the modal dynamic loading and switching controller:

[0104] The status monitor is used to perceive and maintain the global resource status in real time. It periodically polls the utilization rate and operation indicators of each mode (such as latency and packet loss rate) of the heterogeneous public cloud network resource pool through the unified abstract interface of heterogeneous resources, and updates these real-time data (corresponding to the resource capacity vector C and load status in the following text) to the shared memory for the scheduler to query.

[0105] Event handlers, used for coordination: As the signaling entry point of the system, they are responsible for receiving external modal loading requests (MDL requirement files) or internal abnormal alarms triggered by the monitor (such as heartbeat loss). Once an event is received, the event type is parsed and a scheduling request is sent to the scheduler.

[0106] The scheduler, used for decision-making, acts as the core decision engine. Upon receiving a request from an event handler, the scheduler immediately reads the current global resource status from the status monitor, combines it with the requirement constraints in the modal MDL requirement file, and executes a matrix scheduling algorithm or a lossless switching strategy to generate the optimal resource allocation scheme (decision matrix). The command, or switching instruction, is ultimately sent to the modal virtualization layer for execution.

[0107] The lossless switching strategy is as follows: the timing control logic from modal loading request to the completion of lossless switching is as follows: Figure 3 As shown;

[0108] The modal dynamic loading and switching controller first parses the MDL requirement file and generates a dependency graph. After passing resource constraint checks (if not met, a queuing or degradation strategy is triggered), a new modal container is created and the state snapshot of the old modality is restored. Then, it enters the dual-modal parallel stage until the state synchronization is completed, at which point atomic switching is performed. The specific strategy and algorithm logic are as follows:

[0109] S1. Dual-modal state merging: During lossless switching, the system allows the old and new modes to coexist for a short period of time (dual-modal parallelism), within the switching window. Within the modality, the dynamic loading and switching controller issues commands to execute traffic mirroring, simultaneously distributing inbound traffic to both the new and old modalities. During this period, the new modality not only has an initial snapshot state but also processes the mirrored traffic in real time to generate new state increments. The state merging logic is as follows:

[0110]

[0111] This mechanism ensures that incremental data generated during handover (such as newly arrived packets) is not lost. In the formula, This represents the complete state of the new mode at any time t within the switching window; The initial state is calculated using the aforementioned formula; During parallel operation, the state increment generated by the mirrored traffic in the new mode is processed in real time. The dual-modal parallel phase continues until the state difference between the old and new modes converges to a preset threshold (such as state synchronization rate > 99%), and then an atomic switching operation is performed to flip the traffic pointer.

[0112] S2. Switching Window Optimization Constraints: To ensure service quality, the switching process must meet the following optimization objectives:

[0113]

[0114] in, P represents the duration of the window switching. loss This represents the packet loss rate during the handover period. This represents the maximum allowable packet loss rate threshold of the system (e.g.: This constraint ensures that while pursuing rapid switching, the system must prioritize meeting the reliability requirements of the business.

[0115] The modal dynamic loading and switching controller is equipped with a heterogeneous resource elastic scheduling and QoS guarantee mechanism:

[0116] ① Basic resource allocation model: The scheduler aims to maximize the total system utility, and the following optimization model is established:

[0117]

[0118] Among them, u i,k x represents the utility weight of mode i using resource k. i,k This is a Boolean variable, with a value of 1 representing allocation and 0 representing no allocation. It must satisfy the total capacity constraint. d i,k C represents the unit requirement of resource k for running mode i. k Let k be the total physical capacity of resource k.

[0119] This weight supports dynamic updates: the system dynamically adjusts the u of each modality based on real-time monitoring of business SLA achievement rates. i,k The specific adjustment logic is as follows: when monitoring data shows that the SLA indicator (such as latency or packet loss rate) of a certain modality i is close to or worse than a preset threshold, the system automatically increases its corresponding utility weight u. i,k (e.g., execute) , (This is used as a priority boosting factor), thereby increasing the priority of the resources allocated to this mode in the next round of optimization, forcing the scheduler to prioritize the resource needs of high-value or sub-optimal businesses, and triggering online rescheduling;

[0120] ② Vectorized / matrix representation of scheduling problems: such as Figure 4 As shown in the figure, this diagram illustrates the matrix modeling and physical resource mapping principle of elastic scheduling of heterogeneous resources. In actual engineering deployment, in order to use linear programming solvers (such as Gurobi or CPLEX) to quickly solve large-scale nodes, the scheduling problem is vectorized and matrix-modeled. The optimization model belongs to a typical mixed integer linear programming (MILP) problem, which is solved efficiently by commercial solvers. It is assumed that the number of modes to be scheduled in the system is N, and the number of heterogeneous resource types involved is K.

[0121] (21) Define the resource allocation decision matrix :

[0122]

[0123] Among them, matrix elements Let them be Boolean decision variables. , This indicates that the i-th mode has been allocated the k-th type of resource, and 0 indicates that it has not been allocated;

[0124] (22) Define the resource demand moment D and the capacity vector C: Let This is a resource demand matrix, where, Let represent the unit demand of the i-th mode for the k-th type of resource; let This is a resource capacity vector, a column vector representing the total capacity of each of the K types of resources in the system, indicating the upper limit of the physical total amount of each resource.

[0125] (23) Matrix-based constraints: Global resource constraints are uniformly represented as:

[0126]

[0127] in, This represents the Hadamard product of matrices (i.e., element-wise multiplication), a form that directly preserves the modality-resource correspondence and avoids cross-resource interference. This yields an N×K intermediate matrix, where non-zero elements represent the actual amount of resources occupied, and T represents the matrix transpose operation, which transposes the N×K intermediate matrix into a K×N matrix, arranging the occupancy status of the same type of resources in the same row. It is a column vector of dimension N consisting entirely of 1s;

[0128] The engineering physical meaning of this constraint formula is: the system monitors in real time the cumulative occupancy of all currently running modes on various heterogeneous resources such as CPU, GPU, and FPGA, ensuring that it is strictly limited within the total hardware capacity of the physical nodes, thereby fundamentally avoiding system overload or crash due to resource overselling.

[0129] (24) Matrix-based objective function: total system utility U total Represented as:

[0130]

[0131] in, (i.e., on the real number field) The dimensional matrix is ​​a predefined utility weight matrix. It is a dimension of A row vector of all 1s; For dimension A column vector of all 1s; This represents the utility value matrix of the selected allocation schemes. ;

[0132] This formula uses matrix operations to calculate the total utility value of all modes and resource types under the current scheduling scheme in one go, which makes it easy for computers to solve efficiently using the branch and bound method or existing integer programming algorithms.

[0133] ③ Specific Example of Traffic Model and Delay Calculation (Extended Implementation): To verify whether resource allocation meets QoS delay constraints, this implementation provides a specific calculation scenario. Scenario Assumption: A certain "industrial control mode" The data flow passes through the path Includes three links And it goes through two processing nodes. The link and node parameters are as follows:

[0134]

[0135] Node processing latency:

[0136]

[0137] End-to-end delay calculation formula:

[0138]

[0139] Substituting numerical values: To facilitate calculation and prevent confusion of units, the calculation parameters in this embodiment have been normalized: The system defines the standard "Traffic Unit" as the amount of data (i.e., 1 kbit) that takes 1 ms to transmit at a bandwidth of 1 Mbps. The "Queue Length" in the table is measured in this unit, thus the millisecond-level latency can be directly obtained by dividing by the bandwidth value (Mbps).

[0140] a. Queuing and Transmission Delay Calculation: Total Link Delay

[0141] b. Node processing latency calculation: Total node latency

[0142] c. Total end-to-end delay:

[0143] Judgment logic: If the maximum allowable delay of this mode is... ,but The constraints are satisfied, and resource allocation is efficient. If... If the constraint is not met, the scheduler will trigger rescheduling logic (e.g., migrating traffic to a backup path with higher bandwidth, or migrating node processing tasks to a higher-performance FPGA core). The above calculation logic is used to predict whether the modal deployment scheme meets SLA requirements during the scheduling decision phase. If the calculated theoretical latency exceeds the maximum threshold allowed by the service, the scheduler will automatically reject the scheme and trigger rescheduling.

[0144] The system design includes a modal definition and parsing mechanism. To achieve unified definition and resolvable loading of modalities, this system designs a modal description language, which is implemented based on the MDL parser built into the Modal Virtualization Layer (MVL). As the standard input interface for MVL to identify and manage network modalities, the MDL requirement file is usually in JSON or YAML format, covering fields such as protocol stack composition, packet structure, routing policy, QoS constraints, resource requirements, and security policies.

[0145] Specifically, the definition examples and syntax constraints are as follows: MDL requirement files not only support JSON or YAML formats, but also follow the following field constraint syntax structure to ensure the standardization and extensibility of parsing:

[0146] {

[0147] "Modality": <string>,

[0148] "ProtocolStack": <list>,

[0149] "Resources": {"CPU": <string>, "FPGA": <string>, ...},

[0150] "Dependencies": <list>

[0151] }

[0152] Each field supports both static definition and runtime dynamic expansion, ensuring backward compatibility of the parser when handling new modalities (such as future 6G custom protocols).

[0153] Parsing process and dependency graph construction: The modal virtualization layer calls the MDL parser to read the file and generate modal metadata objects in memory. The MDL parser automatically analyzes the Dependencies field and generates a dependency graph—represented by a Directed Acyclic Graph (DAG) structure. It is then merged into the global dependency view. The specific dependency types include: hardware dependencies (such as the need to exclusively use a specific FPGA card), service dependencies (such as the need to rely on the underlying clock synchronization service), and logical dependencies (such as the ICN modality depending on the IPUnderlay modality). If a dependency is missing or physical resources are insufficient, the modal dynamic loading and switching controller will trigger a queuing or resource preemption strategy.

[0154] The microstructure of a single modal container instance and its binding relationship with underlying heterogeneous resources are as follows: Figure 2 As shown, in order to overcome the performance bottleneck of traditional virtualization, the system constructs a deep software-hardware co-mapping mechanism:

[0155] First, the underlying physical network interface card (NIC) uses SR-IOV technology to virtualize multiple virtual functions (VFs) and directly mounts them as virtual interfaces (vNICs) inside the container through a hardware mapping mechanism, achieving zero-copy access to the data plane. Second, the underlying FPGA acceleration card (such as Slot 0) uses PCIe passthrough technology to bypass the host operating system and directly interface with the packet parser inside the container, enabling it to directly call hardware logic for line-speed packet classification and unpacking, significantly reducing processing latency. At the same time, the system uses CPU affinity binding technology to lock the underlying dedicated CPU cores (such as Core 3-4) to serve the protocol stack inside the container, ensuring that computing tasks exclusively use physical cores and avoiding jitter caused by context switching. Thus, an efficient pipelined operation structure is formed inside the container: virtual interface reception - packet parser accelerated processing - protocol stack logic operation - control agent status reporting, achieving near-bare-metal performance while ensuring virtualization isolation.

[0156] The system has built-in robust processing strategies:

[0157] I. Severe Resource Shortage: When the loading of a new modality causes the cumulative resource usage of various types to exceed the threshold, i.e., the following judgment formula is satisfied:

[0158]

[0159] Where F represents the total number of modes currently in operation; This represents the resource requirement k for the new modality to be loaded; C k Given the total physical capacity of resource k, when the above inequality holds, the system executes the following sequentially:

[0160] Ⅰ-1. Global Scan: Locate free nodes for migrated containers;

[0161] Ⅰ-2. Resource Degradation: While ensuring the minimum service level agreement (SLA), temporarily reduce redundant resources in low-priority modalities (such as rolling back GPU acceleration allocated to non-critical modalities to CPU computing) to free up resources for high-priority modalities;

[0162] Ⅰ-3. Waiting in a queue: If the above steps are ineffective, the new modality is placed in the waiting queue;

[0163] II. Multimodal Concurrent Switching Protection: When multiple modes trigger switching simultaneously, the system limits the number of concurrent switches to protect control plane performance. ,in, M is the size of the current concurrent switching set. max The maximum number of concurrent switching allowed by the system is set, and transaction management is performed on batch switching: if any mode switching fails, the state of all related modes in the entire transaction is automatically rolled back;

[0164] Ⅲ. Container Anomaly Self-Healing: The self-healing mechanism is automatically triggered after the state monitor detects an abnormal event. Abnormal events include, but are not limited to: unexpected exit of container process, loss of heartbeat, resource exhaustion (OOM), and deadlock detection. After an anomaly is detected, the system automatically rolls back to the most recent stable state snapshot and restarts the container.

[0165] IV. Adaptive Adjustment of Monitoring Thresholds: To further improve the robustness of the system, an adaptive heartbeat detection mechanism is introduced. The system dynamically adjusts the heartbeat detection cycle and abnormal trigger threshold for modal containers based on the current overall load trend and historical operating data. For example, during periods of high network load and congestion, the timeout judgment threshold is appropriately relaxed to avoid false triggering of the self-healing mechanism due to brief network jitter, thereby achieving a balance between sensitivity and stability.

[0166] The purpose of this embodiment is to provide a multimodal network virtualization and dynamic loading system for public cloud network resources, addressing the shortcomings of existing network virtualization technologies in multimodal environments. Existing technologies primarily focus on fine-grained virtualization of computing, storage, or link resources, making it difficult to achieve overall encapsulation and isolated management at the "network modality" level. Furthermore, in the heterogeneous resource environment of public cloud networks, they lack flexible scheduling and elastic reuse capabilities, and lack mechanisms to support runtime dynamic loading and evolutionary deployment. Therefore, this embodiment proposes an architecture that enables multimodal network symbiosis, independent evolution, and dynamic management on unified physical resources and heterogeneous computing environments.

[0167] This architecture establishes a modal virtualization layer, enabling the encapsulation, isolation, and co-operation of different network modalities within the same physical environment. It also provides a unified modal description and management method, allowing the system to identify, load, switch, and evolve various network modalities. During operation, it can allocate and elastically schedule underlying heterogeneous resources (such as CPUs, GPUs, and FPGAs) on demand, thereby supporting dynamic loading, hot switching, and version rollback of multimodal networks. This architecture is independent of specific network modal types and can adapt to the operational and verification needs of various existing and future network technology systems.

[0168] The above solution overcomes the limitations of traditional network virtualization on single functions or protocol stacks, achieving modal-level virtualization and full lifecycle management. Multimodal networks can be efficiently deployed in public cloud networks and edge computing environments, supporting dynamic switching and evolutionary verification at runtime. This significantly improves resource utilization efficiency and system flexibility, providing fundamental technical support for the construction and verification of multimodal networks in multi-access converged communication systems, future 6G networks, vehicle-to-everything (V2X) networks, industrial internet, and cloud-network convergence environments.

[0169] Example 2:

[0170] This embodiment provides a method for dynamic loading and management of multimodal networks based on modal virtualization, the steps of which are as follows:

[0171] (1) Utilize the heterogeneous public cloud network resource pool to provide feedback on utilization rate and load status;

[0172] (2) Establish a unified abstract model for heterogeneous resources, adopt a standardized mapping mechanism, normalize the physical parameters of heterogeneous public cloud network resource pools into a unified virtual resource descriptor, thereby shielding the differences in underlying hardware, constructing a globally unified virtual resource pool view, then executing resource scheduling algorithms, managing the lifecycle of modalities, and maintaining a global dependency view.

[0173] (3) Maintain global resource status in real time, poll the resource utilization rate and operation indicators (such as latency and packet loss rate) of the heterogeneous public cloud network resource pool, and update these real-time data to the shared memory;

[0174] (4) Receive external events, modal loading requests or internal abnormal alarms triggered by the monitor (such as heartbeat loss). After receiving the event, read the current global resource status, execute the matrix scheduling algorithm or lossless switching strategy, generate the optimal resource allocation scheme or switching instruction, and then issue it for execution.< / list> < / string> < / string> < / list> < / string>

Claims

1. A multimodal network dynamic loading and management system based on modal virtualization, characterized in that, include: The physical resource layer, namely the heterogeneous public cloud network resource pool, serves as the system's bottom layer and includes: general computing resources, accelerated computing resources, programmable logic units, and programmable forwarding chips. It also provides real-time feedback on utilization and load status through a unified abstract interface for heterogeneous resources. The modal virtualization layer is used to establish a unified abstract model for heterogeneous resources, execute resource scheduling algorithms, manage the lifecycle of modalities, and maintain a global dependency view. The application layer is a multimodal network runtime environment, where each modality is encapsulated in an independent virtualization container and supported by a modal container runtime engine. Modal dynamic loading and switching controller: Located above the modal virtualization layer, it is used to issue scheduling instructions to the modal virtualization layer. It consists of a scheduler, a status monitor and an event processor. The three are logically connected through an internal message bus to form a closed-loop control system of perception-decision-execution. The status monitor is used to perceive and maintain the global resource status in real time. It periodically polls the utilization rate and operation indicators of each mode of the heterogeneous public cloud network resource pool through the unified abstract interface of heterogeneous resources, and updates this real-time data to the shared memory for the scheduler to query. Event handlers, used for coordination: As the signaling entry point of the system, they are responsible for receiving external modal loading requests or internal abnormal alarms triggered by the monitor. Once an event is received, the event type is parsed and a scheduling request is sent to the scheduler. The scheduler is used for decision-making. When it receives a request from an event handler, the scheduler will immediately read the current global resource status from the status monitor, combine it with the requirement constraints in the modal MDL requirement file, execute a matrix scheduling algorithm or lossless switching strategy, generate the optimal resource allocation scheme or switching instruction, and finally send it to the modal virtualization layer for execution. The lossless switching strategy is as follows: The modal dynamic loading and switching controller first parses the MDL requirement file and generates a dependency graph. After passing resource constraint checks, it creates a new modal container and restores the state snapshot of the old modality. Then, it enters the bimodal parallel phase until the state synchronization is completed, at which point it performs an atomic switch. The specific strategy and algorithm logic are as follows: S1. Dual-modal state merging: During lossless switching, the system allows the old and new modes to coexist for a short period of time within the switching window. Within the modality, the dynamic loading and switching controller issues commands to execute traffic mirroring, simultaneously distributing inbound traffic to both the new and old modalities. During this period, the new modality not only has an initial snapshot state but also processes the mirrored traffic in real time to generate new state increments. The state merging logic is as follows: ； This mechanism ensures that incremental data generated during the switchover is not lost. In the formula, This represents the complete state of the new mode at any time t within the switching window; This represents the initial running state upon completion of the new mode startup. During parallel operation, the state increment generated by the mirrored flow in the new mode is processed in real time. The dual-modal parallel phase continues until the state difference between the old and new modes converges to a preset threshold, and then an atomic switching operation is performed to flip the flow pointer. S2. Switching Window Optimization Constraints: To ensure service quality, the switching process must meet the following optimization objectives: ； in, P represents the duration of the window switching. loss ϵ represents the packet loss rate during the handover period, and ϵ represents the maximum packet loss rate threshold allowed by the system. This constraint ensures that while pursuing rapid handover, the system must prioritize meeting the reliability requirements of the services.

2. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 1, characterized in that, In the modal virtualization layer: a. Establish a unified abstract model for heterogeneous resources. The modal virtualization layer's built-in resource adapter reads the physical parameters of the physical resource layer. Using a standardized mapping mechanism, the physical parameters are normalized into a unified virtual resource descriptor, and a globally unified virtual resource pool view is constructed. b. Execute resource scheduling algorithm: The modal virtualization layer integrates a scheduling solver interface, collects the remaining capacity status of the physical resource layer in real time, and combines it with the MDL requirement file of the modality to be loaded to transform the scheduling problem into a matrix mathematical model. The modal virtualization layer calls the built-in scheduling solver to calculate the optimal resource allocation decision matrix (X) and generates deployment instructions for specific nodes of the physical resource layer accordingly. c. Managing the lifecycle of a mode: The modal virtualization layer maintains a finite state machine to manage the state transitions of each mode, including: Creation phase: Parse the modal's MDL requirement file, pull the corresponding image or bitstream file, call the container engine to start the modal, and issue hardware configuration instructions; Pause / Migration Phase: Triggers the state snapshot mechanism, saves the context, and freezes the container; Destruction phase: Reclaim computing resources, clean up flow table entries, and update the availability status of the global resource pool; d. Maintain a global dependency view: Based on the dependency graph generated by parsing the modal MDL requirement file, the modal virtualization layer monitors the health status of each dependency in real time to ensure that the prerequisites for modal operation are met.

3. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 2, characterized in that, The state snapshot mechanism works as follows: when a mode is paused or transitioned, a snapshot of the old mode must be saved, and the new mode is initialized using this snapshot. The state transition follows the formula below: ； in, S represents the initial running state when the new mode is started; snapshot This represents the set of state snapshots captured at the switching time t0 of the old mode, specifically including flow tables, buffer queues, session control context, and metadata; S config This represents the basic static configuration of the new modality. The formula indicates that the new modality inherits the complete business context of the old modality, ensuring that existing session connections are not dropped due to reset after the switch.

4. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 3, characterized in that, The modal dynamic loading and switching controller is equipped with a heterogeneous resource elastic scheduling and QoS guarantee mechanism: ① Basic resource allocation model: The scheduler aims to maximize the overall system utility and establishes the following optimization model: ； Among them, u i,k x represents the utility weight of mode i using resource k. i,k This is a Boolean variable, with a value of 1 representing allocation and 0 representing no allocation. It must satisfy the total capacity constraint. d i,k C represents the unit requirement of resource k for running mode i. k Let k be the total physical capacity of resource k. This weight supports dynamic updates: the system dynamically adjusts the u of each modality based on real-time monitoring of business SLA achievement rates. i,k The specific adjustment logic is as follows: when monitoring data shows that the SLA index of a certain modality i is close to or worse than a preset threshold, the system automatically increases its corresponding utility weight u. i,k This will increase the priority of resources allocated to this mode in the next round of optimization, forcing the scheduler to prioritize the resource needs of high-value or sub-optimal businesses and triggering online rescheduling. ② Vectorized / Matrix Representation of the Scheduling Problem: In actual engineering deployments, in order to use linear programming solvers to quickly solve large-scale nodes, the scheduling problem is modeled in a vectorized and matrix manner. The optimization model is a typical mixed-integer linear programming problem, which is solved by a commercial solver. It is assumed that the number of modes to be scheduled in the system is N and the number of heterogeneous resource types involved is K. (21) Define the resource allocation decision matrix : ； Among them, matrix elements Let them be Boolean decision variables. , This indicates that the i-th mode has been allocated the k-th type of resource, and 0 indicates that it has not been allocated; (22) Define the resource demand moment D and the capacity vector C: Let This is a resource demand matrix, where, Let represent the unit demand of the i-th mode for the k-th type of resource; let This is a resource capacity vector, a column vector representing the total capacity of each of the K types of resources in the system, indicating the upper limit of the physical total amount of each resource. (23) Matrix-based constraints: Global resource constraints are uniformly represented as: ； in, Represents the Hadamard product of matrices; This yields an N×K intermediate matrix, where non-zero elements represent the actual amount of resources occupied, and T represents the matrix transpose operation, which transposes the N×K intermediate matrix into a K×N matrix, arranging the occupancy status of the same type of resources in the same row. It is a column vector of dimension N consisting entirely of 1s; (24) Matrix-based objective function: total system utility U total Represented as: ； in, For the preset utility weight matrix, It is a dimension of A row vector of all 1s; For dimension A column vector of all 1s; This represents the utility value matrix that filters out the selected allocation schemes.

5. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 4, characterized in that, The system design modal definition and parsing mechanism relies on the MDL parser built into the modal virtualization layer; This includes defining examples and syntactic constraints; Parsing process and dependency graph construction: The modal virtualization layer calls the MDL parser to read the file and generate modal metadata objects in memory. The MDL parser automatically analyzes and generates a dependency graph—represented by a directed acyclic graph structure—which is then merged into the global dependency view. The specific dependency types include hardware dependencies, service dependencies, and logical dependencies. If a dependency is missing or physical resources are insufficient, the modal dynamic loading and switching controller will trigger a queuing or resource preemption strategy.

6. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 5, characterized in that, To overcome the performance bottlenecks of traditional virtualization, the system constructs a deep software-hardware co-mapping mechanism: First, the underlying physical network interface card (NIC) uses SR-IOV technology to virtualize multiple virtual functions and directly mounts them as virtual interfaces inside the container through a hardware mapping mechanism, achieving zero-copy access to the data plane. Second, the underlying FPGA acceleration card uses PCIe passthrough technology to bypass the host operating system and directly interface with the packet parser inside the container, enabling it to directly call hardware logic for line-speed packet classification and unpacking. At the same time, the system uses CPU affinity binding technology to lock the underlying dedicated CPU core to serve the protocol stack inside the container. Thus, a pipelined operation structure is formed inside the container: virtual interface reception - packet parser accelerated processing - protocol stack logic operation - control agent status reporting.

7. The multimodal network dynamic loading and management system based on modal virtualization as described in claim 6, characterized in that, The system has built-in robust processing strategies: I. Severe Resource Shortage: When the loading of a new modality causes the cumulative resource usage to exceed the threshold, i.e., the following judgment formula is satisfied: ； Where F represents the total number of modes currently in operation; This represents the resource requirement k for the new modality to be loaded; C k Given the total physical capacity of resource k, when the above inequality holds, the system executes the following sequentially: Ⅰ-1. Global Scan: Locate free nodes for migrated containers; Ⅰ-2. Resource Degradation: Temporarily reduce redundant resources in low-priority modes while ensuring the minimum service level agreement is maintained; Ⅰ-3. Waiting in a queue: If the above steps are ineffective, the new modality is placed in the waiting queue; II. Multimodal Concurrent Switching Protection: When multiple modes trigger switching simultaneously, the system limits the number of concurrent switches to protect control plane performance. ,in, M is the size of the current concurrent switching set. max The maximum number of concurrent switching allowed by the system is set, and transaction management is performed on batch switching: if any mode switching fails, the state of all related modes in the entire transaction is automatically rolled back; Ⅲ. Container anomaly self-healing: The self-healing mechanism is automatically triggered after the state monitor detects an abnormal event. Abnormal events include, but are not limited to: unexpected exit of container process, loss of heartbeat, resource exhaustion and deadlock detection. After an anomaly is detected, the system automatically rolls back to the most recent stable state snapshot and restarts the container. IV. Adaptive Adjustment of Monitoring Thresholds: An adaptive heartbeat detection mechanism is introduced. The system dynamically adjusts the heartbeat detection cycle and abnormal trigger threshold for modal containers based on the current overall load trend and historical operating data.

8. A method for dynamic loading and management of multimodal networks based on modal virtualization, applied to the dynamic loading and management system for multimodal networks based on modal virtualization as described in claim 1, characterized in that, The steps are as follows: (1) Utilize the heterogeneous public cloud network resource pool to provide feedback on utilization rate and load status; (2) Establish a unified abstract model for heterogeneous resources, adopt a standardized mapping mechanism, normalize the physical parameters of the heterogeneous public cloud network resource pool into a unified virtual resource descriptor, construct a globally unified virtual resource pool view, then execute the resource scheduling algorithm, manage the life cycle of the modality, and maintain a global dependency view; (3) Maintain the global resource status in real time, poll the resource utilization rate and operation indicators of each mode of the heterogeneous public cloud network resource pool, and update these real-time data to the shared memory; (4) Receive external events, modal loading requests or internal abnormal alarms triggered by the monitor. After receiving the event, read the current global resource status, execute the matrix scheduling algorithm or lossless switching strategy, generate the optimal resource allocation scheme or switching instruction, and then issue it for execution.

Images