Large model information transmission method in heterogeneous network scene

Abstract

The invention discloses a large model information transmission method in a heterogeneous network scene, a heterogeneous network is composed of a non-SDN control domain, an SDN control domain and a large model data center domain, and intra-domain BGP gateways of the large model data center domain, the SDN control domain and the non-SDN control domain are connected by establishing a Peer of a BGP to realize synchronization of large model information; adjacent BGP gateways among the large model data center domain, the SDN control domain and the non-SDN control domain also establish Peer connection of the BGP; the method comprises the following steps of: expanding Notifice of a BGP (Border Gateway Protocol), and transmitting large model information; and perception of all large model information of the whole network is realized. According to the method, the existing network architecture is expanded and compatible by using the BGP, so that the deployment cost is reduced; the SDN controller manages large model information of the whole network in a centralized manner, dynamically decomposes tasks and intelligently matches an optimal model instance, so that the task scheduling efficiency is improved; and the real-time performance of the information is ensured through a periodic or triggered synchronization mechanism.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a large model information transmission method in a heterogeneous network scenario. Background Art

[0002] BGP (Border Gateway Protocol) is an Exterior Gateway Protocol (EGP) used to exchange routing information between autonomous systems (AS). It is primarily used for internet-scale routing management and policy control. While BGP's routing attributes are standardized, users can implement custom attributes to a certain extent through extension mechanisms.

[0003] SDN (Software-Defined Networking) is a new network architecture whose core concept is to separate the control plane from the data plane, allowing unified management and programming of the network through a centralized controller, thereby achieving network flexibility, programmability, and intelligence. In current network architectures, hybrid deployment models, where SDN control domains coexist with traditional non-SDN control domains, have become mainstream.

[0004] For security reasons, in a heterogeneous network environment, TCP / UDP connections cannot be directly established between large models in different large model data center domains, resulting in the inability to transmit and share information between models.

[0005] Due to the dispersed ownership of data assets, in a heterogeneous network environment, large models cannot directly access the original data. They need to obtain the model output results through task calls to achieve cross-domain result sharing. Summary of the Invention

[0006] To this end, the present invention provides a large model information transmission method in a heterogeneous network scenario to solve the problems in the prior art.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A large model information transmission method in a heterogeneous network scenario. The heterogeneous network consists of a non-SDN control domain, an SDN control domain, and a large model data center domain. The large model data center domain contains a large model server and a dedicated BGP gateway; the SDN control domain is composed of multiple BGP gateways and supports SDN centralized control; the non-SDN control domain is composed of traditional BGP gateways and uses a distributed routing protocol.

[0009] The BGP gateways in the big model data center domain, SDN control domain, and non-SDN control domain establish BGP peer connections to synchronize big model information. The adjacent BGP gateways between the big model data center domain, SDN control domain, and non-SDN control domain also establish BGP peer connections to transmit information, realizing the perception of all big model information in the entire network.

[0010] Further: the large model information includes large model capability information, large model task distribution information and large model task result information.

[0011] Further: the large model server deploys the Agent, the Agent collects local large model information, and the Agent establishes a connection relationship with the adjacent gateway.

[0012] Further: establishing a connection relationship between all BGP gateways in the non-SDN control domain; establishing a connection relationship between all BGP gateways in the SDN control domain; establishing a connection relationship between the boundary BGP gateways of the non-SDN control domain and the SDN control domain. After the BGP gateways establish the connection relationship, they exchange large model information to realize the transmission of large model information.

[0013] Further: periodically publish big model information to achieve synchronization of big model information across the entire network; or achieve synchronization of big model information across the entire network through triggered big model information.

[0014] Furthermore: With the help of BGP protocol, a connection relationship is established, and the large model information is transmitted by extending the Notification of the BGP protocol.

[0015] Furthermore: a connection relationship is established between the SDN controller and the adjacent BGP gateway to transmit the large model information, and the SDN controller has all the large model information.

[0016] Furthermore, the SDN controller has all the information of large models and the ability to process large model tasks. The SDN controller receives large model tasks submitted by customers through the northbound API and then executes the following core processes:

[0017] (1) Task decomposition: breaking down customer tasks into multiple subtasks that can be processed in parallel;

[0018] (2) Intelligent matching: dynamically selecting the appropriate model instance based on subtask characteristics and the large model capability library;

[0019] (3) Task distribution: assigning subtasks to the target large model execution nodes through the southbound interface;

[0020] (4) Result aggregation: Collect the subtask results returned by each model, perform consistency verification and integration; feedback closed loop, return the final processing results to the client system through the northbound API.

[0021] Further: Task decomposition and matching based on the capability characteristics of the large model. The capability dimensions include:

[0022] Technical indicator dimensions include throughput speed, response time, and supported modality types;

[0023] The data asset dimension is the original training data set that is unique to the model. This type of data asset cannot be directly called by other models due to ownership or security restrictions, thus forming the core differentiation capability of the model.

[0024] The present invention has the following advantages: using the BGP protocol to expand compatibility with existing network architecture and reduce deployment costs; the SDN controller centrally manages large model information of the entire network, dynamically decomposes tasks and intelligently matches the optimal model instance, improving task scheduling efficiency; and through periodic or triggered synchronization mechanisms, ensures information real-time.

[0025] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] To more intuitively illustrate the prior art and the present application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be considered as limiting conditions for implementing the present application; for example, based on the technical concepts disclosed in this application and the exemplary drawings, those skilled in the art are capable of easily making routine adjustments or further optimizations to the addition / reduction / attribution division, specific shapes, positional relationships, connection methods, dimensional ratios, etc. of certain units (components).

[0027] Figure 1 A network diagram of a large model information transmission method in a heterogeneous network scenario provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following specific embodiments illustrate the implementation of the present invention. People familiar with this technology can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. It should be understood that these embodiments are only to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field can make some non-essential improvements and adjustments to the present invention based on the content of the above invention; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] See also Figure 1 ,A large model information transmission method in heterogeneous network scenarios,

[0030] The heterogeneous network consists of a large-model data center domain, an SDN control domain, and a non-SDN control domain. The large-model data center domain includes a large-model server and a dedicated BGP gateway. The SDN control domain consists of multiple BGP gateways and supports SDN centralized control. The non-SDN control domain consists of traditional BGP gateways and uses a distributed routing protocol.

[0031] like Figure 1 As shown, BGP (Border Gateway Protocol) peer connections (referring to the logical connection between BGP peers) are established between the BGP gateways in all large-model data center domains, SDN control domains, and non-SDN control domains to achieve synchronization of large-model information; information is also transmitted between large-model data center domains, SDN control domains, and non-SDN control domains through BGP peer connections.

[0032] In addition to traditional SDN capabilities, SDN control also has large-scale model information management capabilities, as follows:

[0033] (1) Traditional centralized control capability: the SDN controller has centralized management functions of the SDN network (such as flow table delivery and path calculation);

[0034] (2) Large model management capabilities: maintaining the management of large model information across the entire network and providing full-process support for task processing. The process is as follows: receiving tasks → intelligent decomposition → cross-domain distribution → result aggregation and feedback.

[0035] Example

[0036] The parameters for measuring the performance of large models cover multiple aspects, including model capabilities (such as accuracy, F1 score, and perplexity), efficiency (such as inference time, throughput, and token generation speed), resource consumption (such as memory usage and graphics memory usage), robustness and security (such as anti-attack protection capabilities, bias and fairness), user experience (such as response quality and consistency), and training-related indicators (such as convergence speed and data efficiency).

[0037] To simplify the description, this embodiment uses throughput speed, response time and supported modality types as examples to describe the large model information transmission method. Other parameters (such as model architecture, data characteristics, etc.) can be deduced by analogy.

[0038] By extending the Notification message of the BGP protocol, its Code, Subcode, and Data fields are used to carry information related to the large model. Specifically, it is divided into three categories: (1) large model capability information collection; (2) large model task distribution; (3) large model results; and (4) SDN controller location information.

[0039] The BGP Notification large model capability information collection message carries the large model capability information, as shown in Table 1 below:

[0040] Table 1

[0041]

[0042] After receiving the message, the BGP peer records the information of the large model according to the information in the message and generates a local large model information table. The format of the information table is as follows:

[0043] IP address of the large model Throughput speed Response time Large model mode

[0044] The BGP Notification task distributes messages, carrying the large model's transmission task information, as shown in Table 2 below:

[0045] Table 2

[0046]

[0047] When the SDN controller receives a task from a large model, it sends the task information to the BGP peer through the message until it is sent to the specified large model.

[0048] The BGP Notification result message carries the result information after the large model processing, as shown in Table 3 below:

[0049] Table 3

[0050] Large model information type Code Subcode Data Reply to message 1101 IP address of the large model Reply to message

[0051] When the specified large model is processed, the task information is sent to the BGP peer through the message until it is sent to the SDN controller.

[0052] The Notification SDN controller location message via BGP carries the result information after the large model processing, as shown in Table 4 below:

[0053] Table 4

[0054] Large model information type Code Subcode Data SDN Controller Location 1102 SDN IP address -

[0055] A large model information collection agent is deployed in the large model server (Agent refers to a software entity or program that can run autonomously, perform specific tasks and interact with the large model server environment). The agent has the function of BGP and collects the information of the large model to form the large model information table shown in Table 5 below.

[0056] Table 5

[0057]

[0058] The large model name attribute indicates the name of the large model.

[0059] The IP address attribute indicates the IP address of the proxy server of the large model.

[0060] The MAC address attribute indicates the MAC address of the proxy server of the large model.

[0061] The throughput attribute measures the amount of data or tasks that the model can process per unit time, such as tokens per second.

[0062] The response time attribute represents an indicator that measures the time required for the model to process a request. It usually refers to the total time from the user initiating the request to the model returning the result.

[0063] Large model modality refers to the modality of large models, such as natural language and video.

[0064] The large model information collection agent establishes a BGP peer relationship with the large model data center domain BGP router within the domain, and sends the large model information table to the large model data center domain BGP router through Notification. The large model data center domain BGP router forms a large model information table containing BGP routing information, as shown in Table 6 below:

[0065] Table 6

[0066]

[0067] Router ID attribute, which indicates the unique identifier of the BGP router.

[0068] The large model information is notified to the BGP peer through BGP Notification.

[0069] To further simplify the explanation, we use the response time as an example. The response time of large model A is 1ms, as described below:

[0070] Code: 0x 3E 09; / / code is 1001, indicating the corresponding time attribute;

[0071] Subcode: 0x 02 01 01 01; / / Subcode is the router ID;

[0072] Data: 0x 01; / / represents 1ms, indicating that the response time is 1ms.

[0073] The transmission of large model capability information is carried out through BGP Notification messages, all of which adopt the above-mentioned BGPPeer method; including:

[0074] Large model information transmission between the large model data center domain BGP gateway and the non-SDN control domain BGP gateway;

[0075] Large model information transmission between BGP gateways in non-SDN control domain;

[0076] Large model information transmission between the BGP gateway in the SDN control domain and the BGP gateway in the non-SDN control domain;

[0077] Large model information transmission between the SDN control domain BGP gateway and the SDN controller.

[0078] By transmitting the big model information, the SDN controller and all BGP gateways have the same big model information.

[0079] After the big model information is synchronized, a synchronization timer is used to count. When the synchronization timer reaches the threshold, the big model information collection agent deployed in the big model server initiates the synchronization of all big model information.

[0080] When the information of the large model itself changes, for example, when the response time changes to a certain threshold, an update of the large model information is initiated, and the updated content is delivered incrementally.

[0081] When the SDN controller receives a large model task from a customer, it decomposes the large model task and distributes it to the matching large model.

[0082] The SDN controller has information about all large models, for example:

[0083]

[0084] The SDN controller decomposes the large model tasks as follows:

[0085] Only the SDN controller performs task matching based on the modal information and response time of all current large models. After a successful match, the current task information is sent to the corresponding large model through the large model task message of the Notification message.

[0086] To simplify the explanation, the task decomposition logic is explained using a single task as an example: When the large model task submitted by the customer contains natural language processing capability requirements and low-latency response time indicators, and the data content involves the transportation field (such as "Solutions for low-altitude economy and three-dimensional transportation in the field of intelligent transportation"), the SDN controller determines that the large model C meets the requirements through the capability matching mechanism, and then routes its task to the large model C for execution.

[0087]

[0088]

[0089] Notify the BGP peer of the large model information through BGP Notification.

[0090] The large model sends the result data to the SDN controller via the extended BGP Notification large model result message; after receiving the entry, the controller parses and verifies the result data, integrates the subtask results returned by multiple nodes (such as weighted average or voting), and returns the final result in the user-specified format through the northbound API. The entire process adopts full-link encryption and fault-tolerant mechanisms to ensure the reliability and security of the task closed loop.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A large model information transmission method in a heterogeneous network scenario, characterized in that: It includes a heterogeneous network consisting of a non-SDN control domain, an SDN control domain, and a large-model data center domain. The large-model data center domain includes a large-model server and a dedicated BGP gateway; the SDN control domain is composed of multiple BGP gateways and supports SDN centralized control; the non-SDN control domain is composed of traditional BGP gateways and adopts a distributed routing protocol; The BGP gateways in the big model data center domain, SDN control domain, and non-SDN control domain establish BGP peer connections to synchronize big model information. The adjacent BGP gateways between the big model data center domain, SDN control domain, and non-SDN control domain also establish BGP peer connections to transmit information, realizing the perception of all big model information in the entire network.

2. A large model information transmission method in a heterogeneous network scenario according to claim 1, characterized in that: The large model information includes large model capability information, large model task distribution information and large model task result information.

3. A large model information transmission method in a heterogeneous network scenario according to claim 1, characterized in that: The large model server deploys the Agent, which collects local large model information and establishes a connection with the adjacent gateway.

4. The method for transmitting large model information in a heterogeneous network scenario according to claim 1, characterized in that: For establishing a connection relationship between all BGP gateways in the non-SDN control domain: For establishing a connection relationship between all BGP gateways in the SDN control domain; For establishing a connection relationship between the boundary BGP gateways of the non-SDN control domain and the SDN control domain, after the BGP gateways establish a connection relationship, they exchange large model information to realize the transmission of large model information.

5. The method for transmitting large model information in a heterogeneous network scenario according to claim 1, characterized in that: Periodically publish big model information to synchronize big model information across the entire network; or use triggered big model information to synchronize big model information across the entire network.

6. A large model information transmission method in a heterogeneous network scenario according to claim 1, characterized in that: With the help of BGP protocol, a connection relationship is established, and large model information is transmitted by extending the Notification of BGP protocol.

7. The method for transmitting large model information in a heterogeneous network scenario according to claim 1, characterized in that: The SDN controller establishes a connection relationship with the adjacent BGP gateway to transmit the large model information, and the SDN controller has all the large model information.

8. The method for transmitting large model information in a heterogeneous network scenario according to claim 1, characterized in that: The SDN controller has all the information about large models and the ability to process large model tasks. The SDN controller receives large model tasks submitted by customers through the northbound API and then executes the following core processes: (1) Task decomposition: breaking down customer tasks into multiple subtasks that can be processed in parallel; (2) Intelligent matching: dynamically selecting the appropriate model instance based on subtask characteristics and the large model capability library; (3) Task distribution: assigning subtasks to the target large model execution nodes through the southbound interface; (4) Result aggregation: collect the subtask results returned by each model, perform consistency verification and integration; The feedback loop returns the final processing results to the customer system through the northbound API.

9. The method for transmitting large model information in a heterogeneous network scenario according to claim 7, characterized in that: Task decomposition and matching are performed based on the capability characteristics of the large model. The capability dimensions include: Technical indicator dimensions include throughput speed, response time, and supported modality types; The data asset dimension is the original training data set that is unique to the model. This type of data asset cannot be directly called by other models due to ownership or security restrictions, thus forming the core differentiation capability of the model.

Images