Method for routing reconstruction for distributed model training in hybrid optical-electrical network

Abstract

The application discloses a routing reconstruction method for distributed model training in a hybrid optical-electric network, and belongs to the field of network communication optimization.The application comprises the following steps: based on the current network traffic state and topology information, a cooperative routing model is constructed, and an adaptive weight mechanism is introduced to dynamically adjust the priority of optical channel and electric packet forwarding, thereby realizing efficient cooperative utilization of multi-layer forwarding resources.The method can quickly complete path reconstruction according to the task characteristics and network dynamic state, thereby improving the system resource utilization and reducing the maximum task completion delay.In the large-flow environment generated by large-scale model training, compared with the traditional fixed routing and semi-dynamic routing method, the average completion delay and the maximum delay are improved by more than 45%, and the application has good scalability and real-time performance, and is especially suitable for the burst-flow-intensive intelligent algorithm cluster scene.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of network communication optimization in computer distributed large model training, and particularly relates to a routing reconstruction method for distributed model training in a hybrid optical-electrical network, which is used for dynamically adjusting the routing path in the hybrid optical-electrical network to reduce communication delay. BACKGROUND

[0002] In the current training process of large-scale artificial intelligence models (such as GPT, Gemini, etc.), the model parameter size continues to grow, resulting in high-frequency gradient exchange and parameter synchronization operations between multiple computing nodes. These operations usually involve TB-level data transmission, with huge communication load and strong burstiness and uncertainty, which puts high requirements on the underlying network system. Although the traditional network architecture mainly based on electrical packet switching (EPS) has flexible scheduling capability, it is easy to become a system bottleneck when facing large traffic and low delay communication demands. The hybrid optical-electrical network (HOEN) combines the high bandwidth and low delay characteristics of optical circuit switching (OCS) with the flexible control advantages of electrical packet switching, and is considered as a potential solution to support large model training communication in intelligent computing clusters.

[0003] However, the static or semi-dynamic routing scheme in the current HOEN is difficult to efficiently adjust according to the changing communication mode and network load in the training process, which easily causes low utilization of optical path resources, high task delay and other problems. Therefore, an intelligent routing optimization method with sensing ability, decision-making ability and reconstruction ability is urgently needed to support dynamic routing reconstruction in the training process, so as to effectively control the communication delay and improve the overall system performance. SUMMARY

[0004] The present application aims to solve the problems in the prior art in handling burst communication traffic in distributed large model training, and proposes a routing reconstruction method for distributed model training in a hybrid optical-electrical network, which aims to minimize the maximum communication completion delay.

[0005] To achieve the above-mentioned purpose, the present application proposes a routing reconstruction method for distributed model training in a hybrid optical-electrical network, comprising the following steps:

[0006] According to the parallel mode in the distributed training process of the large model, the traffic characteristics of the burst traffic are analyzed, and the parallel mode includes data parallelism and model parallelism;

[0007] The traffic types are divided into AllReduce traffic and MP traffic based on the traffic characteristics, the AllReduce traffic is adopted in scenarios involving multi-node parameter aggregation, and the communication mode of the MP traffic is determined by a model division strategy;

[0008] The topology and routing are confirmed based on the traffic characteristics based on an optical-electric hybrid network architecture, the optical-electric hybrid network architecture includes an optical layer (an optical network composed of optical links) and an electric layer (an electric network composed of electric links);

[0009] Modeling is performed based on the optical-electric hybrid network architecture, path pre-allocation is performed based on traffic characteristics through the built model, path reconstruction requirements are identified and cross-layer scheduling is performed when traffic dynamically changes, and link load information is integrated in a global range to support subsequent dynamic optimization.

[0010] Further, the topology and routing are confirmed based on the traffic characteristics based on the optical-electric hybrid network architecture, including:

[0011] For different types of communication traffic in model training, the different types of communication traffic are identified and classified according to their distribution and burst intensity, including model parallel traffic, data parallel traffic, and mixed traffic thereof;

[0012] A traffic heat map is formed based on the amount of data transmitted between communication pairs, burst intensity, and path connectivity performance;

[0013] Based on the resource states of optical links and electric links, a multi-path set is constructed in combination with the traffic characteristics and communication requirements (amount of data to be transmitted) of inter-task communication pairs;

[0014] A path grading strategy is adopted to preferentially map high-intensity burst traffic to optical links, and the remaining traffic is mapped to electric links, thereby realizing heterogeneous allocation of paths;

[0015] The communication topology structure between nodes is dynamically adjusted according to the current network load state and historical burst trends, and the optical link reconstruction is triggered if necessary;

[0016] The selected routing paths are monitored and fed back in real time, and if congestion or bottleneck traffic is detected, a path reconstruction mechanism is triggered to perform local or global path update, so as to guarantee controllability of communication delay and system load balancing.

[0017] Further, the modeling based on the optical-electric hybrid network architecture includes defining an optimization problem and parameter settings based on the optical-electric hybrid network architecture;

[0018] Constraint conditions in the optimization process are defined, including reconstruction compatibility, optical path continuity, link capacity constraints, and traffic schedulability.

[0019] Further, the definition optimization problem is specifically: under the conditions of given communication demand, link capacity, transmission rate and reconstruction overhead, jointly optimizing routing and path allocation to minimize the maximum communication delay, and meeting the topology and resource constraints.

[0020] Further, the path pre-allocation based on traffic characteristics through the built model comprises:

[0021] According to the traffic demand matrix of the communication pair, the communication flow is sorted in descending order of total data volume, and the communication pair with larger transmission data volume is preferentially allocated with a path;

[0022] For each communication pair, a plurality of feasible paths are searched as candidate paths, the total communication delay of each candidate path is calculated, and the candidate path with the minimum delay is selected as the optimal path for allocation;

[0023] After each path allocation is completed, the link utilization rate is updated in real time to dynamically reflect the network load state.

[0024] Further, the path reconstruction demand is identified and cross-layer scheduling is performed when the traffic dynamically changes, comprising:

[0025] Under the dynamic traffic environment, by comparing the path allocation results of the current time and the last time, the communication pair set that needs to be allocated with a path is identified;

[0026] For the reconstruction traffic, temporary path dredging is preferentially performed in the electrical layer, and an electrical layer temporary path dredging algorithm is called to perform electrical layer path allocation;

[0027] If the communication delay of the electrical layer temporary path is higher than the reconstruction time of the optical path, the remaining traffic is migrated to the optical layer to complete the transmission;

[0028] The electrical layer dredging delay and the optical layer transmission delay jointly constitute the total communication delay.

[0029] Further, the link load information is integrated in the global range to support subsequent dynamic optimization, comprising: in the path updating process, the load information of the electrical layer temporary path and the optical layer path is uniformly integrated to form a global link load view, which provides an optimization basis for subsequent dynamic routing scheduling and resource allocation.

[0030] The application also provides a routing reconstruction device for distributed model training in a hybrid optical-electrical network, comprising one or more processors for implementing the routing reconstruction method for distributed model training in the hybrid optical-electrical network.

[0031] The application further provides an electronic device, comprising a memory and a processor, the memory being coupled with the processor; wherein the memory is used for storing program data, and the processor is used for executing the program data to realize the method for reconstructing routing in a hybrid optical-electric network facing distributed model training.

[0032] The application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the method for reconstructing routing in a hybrid optical-electric network facing distributed model training.

[0033] The application has the beneficial effects that the application proposes a dynamic optical-electric collaborative routing reconstruction algorithm for coping with burst traffic in distributed training, aims to minimize the maximum communication completion delay, and the core innovation lies in constructing a collaborative optimization framework of optical layer and electric layer. Firstly, the application fully considers key constraints such as transmission demand constraint, optical link transmission condition constraint, reconstruction occurrence condition and link capacity, and models the reconstruction scheme of communication pairs. Secondly, a dynamic reconstruction algorithm based on minimizing the maximum communication delay is proposed, and the path decision in the burst traffic scenario is converted into an extreme optimization problem of minimizing the maximum delay. The core process of the algorithm is divided into two stages: 1) temporary relief stage of the electric layer: when it is detected that the burst traffic causes the performance of the original optical path to deteriorate, the affected traffic is immediately switched to the electric layer path forwarding, and the temporary optimal path is allocated to the affected traffic based on the minimum maximum allocation algorithm. The goal of this stage is to ensure that the communication traffic in distributed training is not affected. 2) optical layer reconstruction optimization stage: the optical link reconfiguration process is triggered synchronously, and a delay-aware resource allocation mechanism is used to globally optimize and schedule the optical link. After the optical path reconstruction is completed, the communication flow is seamlessly migrated back to the optical layer, so as to eliminate the long-term delay accumulation by using the large bandwidth characteristics of the optical circuit, and finally realize the goal of minimizing the maximum communication delay. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 The optical-electric hybrid network architecture diagram of the method of the application;

[0036] Figure 2 The parallel mode schematic diagram of the application; wherein (a) is a data parallel schematic diagram, and (b) is a model parallel schematic diagram;

[0037] Figure 3Fig. 1 is a communication flow type heat map of the present application; wherein (a) is an MP flow heat map, and (b) is an AllReduce flow heat map;

[0038] Figure 4 Fig. 4 is a flow chart of an optimal path distribution algorithm based on minimizing maximum delay of the present application;

[0039] Figure 5 Fig. 5 is a flow chart of a dynamic routing reconstruction algorithm of the present application;

[0040] Figure 6 Fig. 6 is a comparison chart of MP flow maximum communication delay trends in a burst flow scenario provided by an embodiment of the present application;

[0041] Figure 7 Fig. 7 is a comparison chart of AllReduce flow maximum communication delay trends in a burst flow scenario provided by an embodiment of the present application;

[0042] Figure 8 Fig. 8 is a comparison chart of Hibrid flow maximum communication delay trends in a burst flow scenario provided by an embodiment of the present application. DETAILED DESCRIPTION

[0043] The present application will be described in detail below with reference to the accompanying drawings. The features in the following embodiments and implementation manners can be combined with each other without conflict.

[0044] The present application performs intelligent topology reconstruction and routing optimization of an optical-electric hybrid network under the premise of burst flow occurring in the training process in the scenario of distributed large model training, so as to realize the improvement of resource utilization and the minimization of maximum task completion delay.

[0045] The specific implementation steps are as follows:

[0046] Step 1, first analyze the flow mode according to the parallel mode of distributed training (such as shown in FIG. 2). Analyze the flow characteristics according to the existing data parallel and model parallel two parallel modes. Figure 2

[0047] Step 2, divide the flow type into AllReduce flow and MP flow according to the flow characteristic analysis (such as shown in FIG. 3), the communication mode of MP flow is strongly related to the model division strategy, and AllReduce flow is required in the scene involving multi-node parameter aggregation regardless of data or model parallel. Figure 3

[0048] Step 3, design the input of the optical-electric hybrid network according to the large transmission capacity of the optical link and the flexibility of the electrical link (such as shown in FIG. 4). Figure 1 ​​The illustrated optoelectronic hybrid network architecture includes an optical network architecture and an electrical network architecture. The electrical network part is a typical Fat-tree architecture, which is divided into core layer, aggregation layer and access layer. Each switch in a k-ary Fat-tree has k ports. The core layer is the top layer, which has a total of k switches, and has k pods, each of which has k switches. The aggregation layer and the access layer each have k switches. Each switch in the access layer can accommodate k servers, so all the pods can accommodate k servers. The optical network part is composed of k optical switches and servers, and is a shared direct connection structure. Each switch has k interfaces connected to k optical switches. The optoelectronic hybrid network is modeled as a network topology.

[0049] Step 4, modeling according to the optoelectronic hybrid network architecture.

[0050] Step 5, based on the above modeling, a cross-layer joint routing optimization algorithm for the optoelectronic hybrid network is designed. The algorithm is divided into three main stages: first, path pre-allocation based on traffic characteristics; then, when the traffic changes dynamically, identify the path reconstruction requirement and perform cross-layer scheduling; finally, integrate link load information in the global range to support subsequent dynamic optimization.

[0051] In one embodiment, step 4 further includes the following steps:

[0052] (4.1) Based on the optoelectronic hybrid architecture, the problem is formalized as follows: under the conditions of given communication demand, link capacity, transmission rate and reconstruction overhead, jointly optimize routing and path allocation to minimize the maximum communication delay and meet the topology and resource constraints.

[0053] (4.2) Define the key parameters as shown in Table 1 below:

[0054] Table 1

[0055] (4.3) When jointly optimizing the path allocation and routing optimization of the optoelectronic hybrid network architecture, the constraint conditions are defined as follows:

[0056] Reconstruction compatibility: if the link on any optical path from node to node changes before and after reconstruction, it means that reconstruction has occurred. That is:

[0057] .

[0058] ​​​​​​​​Lightpath continuity: If a lightpath is reconfigured, the reconfigured lightpath must pass through at least one optical switch to ensure the continuity of the lightpath, i.e.:

[0059] ;

[0060] .

[0061] Link capacity constraint: The load of an optical / electrical link is limited by its capacity, i.e.:

[0062] ;

[0063] .

[0064] Traffic schedulability: The capacity of an electrical lightpath must satisfy the traffic demand during reconfiguration to ensure that burst traffic can be temporarily carried by the electrical layer, i.e.:

[0065] .

[0066] (4.4) Taking the maximum communication delay as the optimization objective, i.e.:

[0067] .

[0068] In one embodiment, the step 5 further comprises the following steps:

[0069] (5.1) Designing an optimal path allocation algorithm based on minimizing the maximum delay, performing traffic classification and path pre-allocation, and the algorithm flow chart is shown in Figure 4 .

[0070] First, the communication flows are sorted in descending order according to the traffic demand matrix of the communication pairs, and the paths are allocated preferentially to the communication pairs with larger transmission data volume, so as to avoid the problem that large traffic communication occupies too many resources and causes small traffic communication to be unable to be completed in time.

[0071] Subsequently, k feasible paths are searched for each communication pair as candidate paths, the total communication delay of each candidate path is calculated, and the candidate path with the minimum delay is selected as the optimal path for allocation.

[0072] After each path allocation is completed, the link utilization rate is updated in real time to dynamically reflect the network load state and provide accurate feedback for the scheduling of subsequent traffic, so as to avoid local congestion problems.

[0073] (5.2) Designing a dynamic routing reconfiguration algorithm, making a dynamic topology reconfiguration decision, and the algorithm flow chart is shown in Figure 5 .

[0074] In the dynamic flow environment, first, by comparing the path allocation results of the current time and the last time, the communication pair set that needs to be re-allocated is identified.

[0075] For the reconstruction flow, the temporary path dredging in the electrical layer is preferentially attempted, and the electrical layer temporary path dredging algorithm is called to perform electrical layer path allocation.

[0076] If the communication delay of the electrical layer temporary path is higher than the reconstruction time of the optical path, the remaining flow is migrated to the optical layer to complete transmission. The final total communication delay is jointly determined by the electrical layer dredging delay and the optical layer transmission delay, which embodies the synergistic effect of the cross-layer scheduling mechanism.

[0077] (5.3) Path update and global link state integration. During the path update process, the load information of the electrical layer temporary path and the optical layer path is uniformly integrated to form a global link load view, which provides an optimization basis for subsequent dynamic routing scheduling and resource allocation, and further improves the efficiency and corresponding ability of collaborative utilization of optical and electrical resources.

[0078] In one embodiment, the method provided by the present application is designed for comparative experiments, and the experimental parameter algorithm program is written using Python 3.8. The experiment is run on a server configured with an Intel Core i9-14900K processor and 128GB of memory, and the experiment scale is a hundred-kilocal cluster scale. Other experimental environment parameters are set as follows: the optical switch reconstruction delay is 10ms, the switch processing delay is 1us, the optical link capacity is 20Gbps, and the electrical link capacity is 10Gbps. The link configuration is set in a scaled manner, aiming to build a controllable simulation environment to reflect the influence of different link capabilities in the optical and electrical hybrid network on the response capability of the scheduling strategy, especially in the distributed training scenario where burst traffic frequently occurs. The optical link bandwidth is set to twice the electrical link bandwidth, which can not only reflect the performance advantage of the optical link in forwarding large traffic, but also avoid interfering with the performance evaluation of the scheduling algorithm itself due to excessive magnification of the gap.

[0079] As shown in Figure 6-8 The algorithm program provided by the present application has designed and implemented two representative routing algorithms for comparative experiments according to existing research, which are a fixed routing algorithm based on a fixed strategy (Hybrid Static) and a semi-update algorithm based on a semi-update mechanism (Semi Update), for evaluating the performance difference in the burst traffic scenario.

[0080] Compared with the traditional fixed routing and semi-dynamic routing methods, the method provided by the present application can improve the maximum delay by more than 45% in the large-flow environment generated by large-scale model training, has good scalability and real-time performance, and is especially suitable for burst traffic-intensive intelligent computing cluster scenarios.

[0081] The embodiment of the present application also provides a device for route reconstruction for distributed model training in a hybrid optical-electric network, comprising one or more processors, which are used to implement the method for route reconstruction for distributed model training in a hybrid optical-electric network.

[0082] The implementation process of the functions and roles of the units in the device is specifically shown in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0083] For the device embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present application scheme. Those skilled in the art can understand and implement it without creative labor.

[0084] The embodiment of the present application also provides an electronic device, comprising a memory and a processor, wherein the memory is coupled with the processor; the memory is used to store program data, and the processor is used to execute the program data to implement the method for route reconstruction for distributed model training in a hybrid optical-electric network according to any embodiment.

[0085] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the method for route reconstruction for distributed model training in a hybrid optical-electric network.

[0086] The computer readable storage medium can be an internal storage unit of any device with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit of any device with data processing capability and an external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the device with data processing capability, and can also be used to temporarily store data that has been output or will be output.

[0087] The above merely describes preferred embodiments of the present application, but is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0088] The above embodiments are only used to illustrate the design ideas and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the present application and implement it, and the protection scope of the present application is not limited to the above embodiments. Therefore, any equivalent change or modification made according to the disclosed principles and design ideas of the present application is within the protection scope of the present application.

Claims

1. A method for route reconfiguration for distributed model training in a hybrid optical-electrical network, characterized in that, The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend.

2. The method of claim 1, wherein, The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend.

3. The method of claim 1, wherein, The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend.

4. The method of claim 3, wherein, The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend.

5. The method of claim 1, wherein, The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to a method for dynamically adjusting a communication topology structure among nodes based on a current network load state and a historical burst trend. The application relates to The communication flows are sorted in descending order of total data volume according to a traffic demand matrix of the communication pairs, and paths are allocated preferentially to communication pairs with larger data volume; For each communication pair, a number of feasible paths are searched as candidate paths, total communication delay of each candidate path is calculated, and the candidate path with the minimum delay is selected as the optimal path for allocation; After each path allocation, link utilization is updated in real time to dynamically reflect network load status.

6. The method of claim 1, wherein, The global link load information integration to support subsequent dynamic optimization includes: during path updating, load information of the electrical layer temporary path and the optical layer path is integrated to form a global link load view, which provides an optimization basis for subsequent dynamic routing scheduling and resource allocation.

7. An apparatus for routing reconstruction oriented to distributed model training in a hybrid optical-electrical network, characterized in that, The one or more processors are configured to implement the method for route reconstruction for distributed model training in a hybrid optical-electrical network according to any one of claims 1-6.

8. An electronic device comprising a memory and a processor, characterized in that The memory is coupled to the processor, and the memory is configured to store program data, and the processor is configured to execute the program data to implement the method for route reconstruction for distributed model training in a hybrid optical-electrical network according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the method for route reconstruction for distributed model training in a hybrid optical-electrical network according to any one of claims 1-6.

Images