Optical circuit switch reconfiguration scheduling training acceleration system and method based on maximum network flow

Abstract

The application provides an optical circuit switch reconstruction scheduling training acceleration system and method based on a maximum network flow. A heuristic scheme is used to optimize the parallel dimension of distributed training and the grouping of training tasks, reduce the communication flow across an optical circuit switch (OCS), and then, under the optimal grouping, select the optimal OCS connection configuration at the corresponding moment based on a maximum network flow algorithm to accelerate the global training time. By using a heuristic pruning method to optimize the combination relationship of data parallelism, tensor parallelism and pipeline parallelism, and placing the training tasks according to the rack where the GPU is located and the OCS domain, the situation of high-frequency tensor parallel communication across the OCS can be reduced.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically to a system and method for accelerating training based on maximum network flow optical circuit switch reconfiguration scheduling. More particularly, it relates to a method for accelerating distributed large language model training based on maximum network flow and end-to-end intelligent computing center optical circuit switch reconfiguration scheduling. Background Technology

[0002] Currently, a common approach in traditional optical circuit-switched intelligent computing center designs is to use a fixed polling OCS scheduling strategy. While this can adapt to high traffic loads and provide dynamism, the OCS connection relationships are irrelevant to the current training traffic demands and cannot identify bursts of communication during large language model training (see the February 2020 conference "17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20)" which published "Expanding across time to deliver bandwidth efficiency and low latency"). Traditional OCS scheduling schemes optimize OCS connections and scheduling for collective communication by constructing a cost function, but this approach is complex and only considers the collective communication step in distributed training, lacking consideration for computation and modeling the overlap between computation and communication (see the May 2025 issue of the journal "Journal of Optical Communications and Networking" pp. 470-484, which published "RECCL: Optimizing collective algorithms for reconfigurable optical networks"). Traditional OCS scheduling schemes for accelerating distributed training, while considering the dynamic adjustment of OCS optical path interconnect reconfiguration according to the training phase, lack consideration of optimizing the placement of training tasks based on communication volume to optimize the total communication overhead during training (see “Training-phase-aware optical circuit switching reconfiguration for large language model” disclosed in M3F.5 of the “Optical Fiber Communication Conference (OFC2026)” in March 2026).

[0003] Patent application CN120769195A discloses a method, apparatus, electronic device, medium, and computer program product for optical path reconstruction of network topology. The method includes determining a current optical path scheme and determining the communication and load information of the target network based on the current topology of the target network; constructing an objective function that varies with the target optical path scheme based on the current optical path scheme, communication information, and load information, and determining an optimized target optical path scheme with the goal of minimizing the objective function; modifying the current topology of the target network based on the optimized target optical path scheme to obtain a reconstructed topology of the target network, which is used as the basis for processing data transmission tasks. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of this invention. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a system and method for accelerating optical circuit switch reconfiguration scheduling training based on maximum network flow.

[0005] The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow provided by the present invention includes: Workload generation module: Used to generate training workload for large language models based on the input large language model framework, parallel dimension, number of parameters and GPU type parameters; Grouping module: connected to the workload generation module, used to group all GPUs participating in training in the intelligent computing center according to the generated training workload, to obtain data parallel group, tensor parallel group and pipeline parallel group, and to determine the computation and communication tasks within and between groups and the corresponding GPU groups; Directed Acyclic Graph Construction Module: Connected to the grouping module, it is used to create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups. The directed acyclic graph is used to model the order and dependency of computation and communication tasks. Heuristic pruning module: connected to the directed acyclic graph construction module, used to reduce the corresponding parallel dimension and GPU grouping of the directed acyclic graph in the training process with relatively dense cross-optical circuit switch communication traffic through heuristic pruning algorithm, and select the optimal GPU grouping and computation task allocation results; Training execution and traffic monitoring module: connected to the heuristic pruning module, used to execute the complete training computation task and communication task according to the selected optimal GPU group and computation task allocation result, in accordance with the order and dependency relationship of the directed acyclic graph, and generate a traffic matrix of cross-optical circuit switch communication based on the communication of each group set; Optical circuit switch scheduler: connected to the training execution and traffic monitoring module and the optical circuit switch respectively, used to select the optimal optical circuit switch interconnection topology for the corresponding time slot according to the traffic matrix of cross-optical circuit switch communication, and issue the corresponding route, and execute the communication of each group of training tasks under the interconnection topology.

[0006] Preferably, the workload includes the overhead of communication categories, communication data volume, and computation time required for each layer of the specific model. The overhead is determined by the type of input GPU, the parallel dimension of model training, the structure of the model, the training framework of the model, the number of parameters of the model, and the parameters of each dimension of the model.

[0007] Preferably, the intra-group and inter-group computation and communication tasks and the corresponding GPU groups are determined according to the framework and parallel dimension of the training process of the large language model. When the execution reaches the corresponding layer, the corresponding computation or communication overhead is written into the workload according to the GPU position of the layer, the input parameters and the parallel dimension. The grouping results are used to determine the mapping relationship between the GPUs participating in each training sub-process in the actual intelligent computing center and the workload.

[0008] Preferably, the directed acyclic graph of the training process is modeled according to the GPU grouping results and the forward and backward propagation of the training. Each node of the directed acyclic graph represents the computation or set communication of the corresponding step performed by the corresponding GPU. The edges of the directed acyclic graph represent the dependency relationship between the preceding steps and the subsequent steps. Each node needs to wait for all preceding steps to be completed before it can start execution.

[0009] Preferably, the objective function of the heuristic pruning algorithm is expressed as:

[0010] in, This represents the communication volume between GPU i and GPU j. This indicates whether cross-OCS communication is required between GPU i and GPU j.

[0011] The heuristic pruning algorithm avoids cross-optical circuit switch communication by restricting communication-intensive tensor parallel groups to the same rack and reducing cross-domain traffic by allocating data parallel groups to the same optical circuit switch domain. The algorithm sorts candidate schemes according to the heuristic pruning algorithm, removes schemes whose communication cost ranking does not meet the preset requirements, and retains the best or multiple schemes that meet the preset requirements to enter the subsequent training execution and optical circuit switch scheduling optimization process.

[0012] Preferably, the traffic matrix of the cross-optical circuit switch communication originates from the selected communication events of a directed acyclic graph obtained after optimization by a heuristic algorithm. Each communication event includes information on the type of communication, the GPU groups participating in the communication, and the amount of data communicated.

[0013] Preferably, the maximum network flow algorithm is an improved Dinic maximum network flow algorithm, and its expression is:

[0014] in, This is the current scheduling slot; It serves as a super source, used to aggregate all sending requests; For a certain source node; To expand the graph edge set, including source edges, destination edges, and candidate optical path edges; Indicates time slot From super source node to source node The allocated traffic.

[0015] Preferably, the process of generating the maximum network flow matching optical circuit switch interconnection topology for the corresponding time slot is as follows: In time slot t, the cross-optical circuit switch communication requirements of the current ready training task are extracted and aggregated into a demand graph. Then, an extended flow graph is constructed by combining the port, plane and bandwidth constraints of the optical circuit switches. Subsequently, the Dinic algorithm is used to solve the maximum flow, and the flow is sorted according to the traffic size of the source node and the destination node pair. High-demand connections are selected first to generate the cost time slot optical circuit switch interconnection topology, and the corresponding routes are issued accordingly.

[0016] Preferably, after entering the next scheduling time slot, the optical circuit switch scheduler re-extracts the traffic matrix of cross-optical circuit switch communication and repeatedly executes the operation of selecting the optimal optical circuit switch interconnection topology for the corresponding time slot through the maximum network flow algorithm until the entire training process is completed.

[0017] The optical circuit switch reconfiguration scheduling training acceleration method based on maximum network flow provided by the present invention includes the following steps: Step 1: Given the corresponding large language model framework, parallel dimension, parameter size, and GPU type parameters, generate the training workload of the large language model; Step 2: Based on the current training workload, group all GPUs participating in training in the intelligent computing center into data parallel, tensor parallel, and pipeline parallel groups, thereby obtaining different intra-group and inter-group computing and communication tasks and corresponding GPU groups; Step 3: Create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups, which is used to model the sequence and dependencies of computation and communication tasks; Step 4: Use a heuristic pruning algorithm to remove the corresponding parallel dimension and GPU grouping of the directed acyclic graph in the training process where the communication traffic across the optical circuit switch is relatively dense, and select the optimal GPU grouping and computation task allocation results. Step 5: Based on the selected optimal GPU grouping and computation task allocation results, execute the complete training computation task and communication task according to the order and dependency of the directed acyclic graph. In this process, a traffic matrix of cross-optical circuit switch communication based on the communication of each group set is generated. Step 6: Based on the traffic matrix of cross-optical circuit switch communication, generate the maximum network flow matching optical circuit switch interconnection topology for the corresponding time slot according to the improved Dinic maximum network flow algorithm, and execute the communication of each group of training tasks under this interconnection topology; repeat step 6 after entering the next scheduling time slot to complete the entire training process.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) By adopting a heuristic pruning method to optimize the combination relationship of data parallelism, tensor parallelism and pipeline parallelism, and placing training tasks according to the rack and OCS domain of the GPU, the situation of high-frequency tensor parallel communication crossing OCS can be reduced. This method solves the problems of high-frequency ensemble communication crossing rack transmission, excessive cross-OCS communication volume and unbalanced OCS link utilization that are easy to occur in the existing training task layout. It can reduce training communication overhead and improve efficiency.

[0019] (2) By adopting the maximum network flow algorithm, the cross-OCS aggregate communication generated during the training process is modeled as a multi-source and multi-destination maximum flow problem. The OCS connection relationship can be selected according to the actual communication needs of the current time slot. Compared with the fixed polling OCS topology, the present invention can enable the OCS connection to prioritize the source node-destination node pair with high traffic and high urgency, and solve the problems of existing OCS reconstruction strategies being insensitive to large language model training traffic and the mismatch between optical path resources and training communication needs.

[0020] (3) The design of this method takes into account the different communication modes generated by data parallelism, tensor parallelism and pipeline parallelism. It can adapt to the complex characteristics of overlapping computing tasks and communication tasks, bursty aggregate communication, and cross-rack traffic changes with the training stage in the training of large language models. This method is not only applicable to single model scale, but also to optical interconnected intelligent computing center training scenarios with different parameter scales, different sequence lengths, different number of GPUs and different number of OCS planes. Attached Figure Description

[0021] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1A schematic diagram illustrating the entire process of scheduling optimization for distributed training optical circuit switches in intelligent computing centers; Figure 2a This is a schematic diagram of a heuristic pruning algorithm. Figure 2b The graph shows a reconstructed scheduling algorithm for multi-point to multi-point maximum network flow in aggregate communication. Figure 3 The performance test results are shown in the graph (heuristic pruning optimization results). Figure 4 The graph shows the performance test results (training time per iteration). Figure 5 This is a graph showing the performance test results (All-Gather flow completion time). Detailed Implementation

[0022] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0023] Example like Figure 1 This invention provides a method for accelerating the training of distributed large language models by reconfiguring and scheduling optical circuit switches in a full-process intelligent computing center based on maximum network flow, comprising the following steps: Step 1: Generate training workload: Given the corresponding large language model framework, parallel dimension, number of parameters, type of GPU, etc., generate the training workload of the large language model. Step 2: Group the GPUs participating in training: Based on the current training workload, group all the GPUs participating in training in the intelligent computing center into data parallel, tensor parallel, and pipeline parallel groups, thereby obtaining different intra-group and inter-group computing and communication tasks and corresponding GPU groups. Step 3: Establish a directed acyclic graph of the entire computation and communication tasks: Create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups. This graph is used to model the sequence and dependencies of computation and communication tasks and to simulate the overlap between computation and communication. Step 4, Heuristic Pruning Algorithm: The heuristic pruning algorithm is used to remove the corresponding parallel dimensions and GPU groups of the directed acyclic graph in the training process with relatively dense cross-OCS communication traffic, and select the optimal GPU group and computation task allocation results. like Figure 2a The goal of heuristic pruning is to minimize cross-OCS communication traffic while ensuring load balancing between computation and communication. The objective function can be expressed as follows: ,in, This represents the communication volume between GPU i and GPU j. This indicates whether cross-OCS communication is required between GPU i and GPU j. Specifically, communication-intensive tensor parallel groups are confined to the same rack to avoid cross-OCS communication. For data parallel groups, due to the large gradient synchronization traffic, DP groups are preferentially assigned to the same OCS domain to reduce cross-domain traffic.

[0024] Step 5: Execute the complete training process according to the directed acyclic graph (DAG): Based on the selected optimal GPU grouping and computation task allocation results, execute the complete training computation and communication tasks according to the order and dependencies of the DAG. During this process, a traffic matrix for cross-OCS communication based on the communication of each group is generated. The scheduler executes computation and communication concurrently in batches according to the readiness relationship of the DAG. For each batch of cross-OCS communication, the demand is first extracted and constructed into a multi-source, multi-sink maximum flow problem. Then, communication traffic is injected into the unified optical interconnect simulation network, and latency is estimated using an alpha-beta model, ultimately obtaining the end-to-end training iteration time. For example... Figure 4 The graph shows the performance test results for a single iteration of training.

[0025] Step 6: Select the optimal OCS interconnection topology using the maximum network flow algorithm and complete the traffic transmission: (e.g., ...) Figure 2b Based on the cross-OCS communication traffic matrix, the improved Dinic maximum network flow algorithm generates the maximum network flow matching OCS interconnection topology for the corresponding time slot, and performs communication for each training task under this interconnection topology. The Dinic maximum network flow expression can be represented as: ,in For the current scheduling slot, As a super source, it is used to aggregate all sending requests. For a certain source node, To expand the graph edge set to include source edges, destination edges, and candidate optical path edges, Indicates time slot From super source node to source node The allocated traffic. The complete process of generating the maximum network flow matching OCS interconnection topology for the corresponding time slot is as follows: In the time slot, the cross-OCS communication requirements of the current ready training task are first extracted and aggregated into a demand graph. Then, an extended flow graph is constructed by combining OCS port, plane, and bandwidth constraints. Subsequently, the Dinic algorithm is used to solve for the maximum flow, sorted by the traffic size on the source node and destination node pairs, and high-demand connections are selected first to generate the OCS interconnection topology for the current time slot, and the corresponding routes are issued accordingly.

[0026] After entering the next scheduling slot, repeat step six to complete the entire training process.

[0027] In step one, the training workload includes the type of set communication, the amount of communication data, and the computation time overhead required for each layer of the specific model. These specific overheads are jointly determined by the type of input GPU, the parallel dimension of model training, the structure of the model, the training framework of the model, the number of parameters of the model, and the parameters of each dimension of the model.

[0028] In step two, the intra-group and inter-group computation and communication tasks and the corresponding GPU groupings are determined according to the framework and parallel dimension of the training process of the large language model. When the execution reaches the corresponding layer, the corresponding computation or communication overhead is written into the workload according to the GPU position of the layer, the parameters input in step one, and the parallel dimension. The grouping results in this step determine the mapping relationship between the GPUs participating in each training sub-process in the actual intelligent computing center and the workload.

[0029] In step three, the directed acyclic graph of the training process is modeled according to the GPU grouping results of the previous step, following the forward and backward propagation of training. Each node of the directed acyclic graph represents the computation or collective communication performed by the corresponding GPU in the corresponding step. The edges of the directed acyclic graph represent the dependencies between the preceding and subsequent steps. Each node needs to wait for all preceding steps to be completed before it can start execution. This can model the overlap of computation and communication tasks in the training process of a large language model.

[0030] In step four, the heuristic pruning algorithm is implemented by reducing cross-OCS communication traffic. The tensor parallelism dimension is constrained to the number of GPUs available within the same rack, while the data parallelism dimension is selected based on minimizing the total cross-OCS communication traffic. According to this heuristic pruning algorithm, candidate schemes are ranked, and schemes with lower communication costs are pruned, retaining the optimal scheme or several relatively optimal schemes for subsequent training execution and OCS scheduling optimization processes. Figure 3 The image shows the results of heuristic pruning optimization.

[0031] In step five, the cross-OCS communication traffic matrix originates from selected communication events in a directed acyclic graph (DAG) optimized using a heuristic algorithm. Each communication event includes information such as the type of communication (e.g., All-Reduce, All-Gather), the GPU groups participating in the communication, and the amount of data communicated. The cross-OCS communication traffic matrix is ​​generated based on this information. Figure 5 The graph shows the performance test results for the completion time of the All-Gather flow.

[0032] In step six, the improved maximum network flow algorithm is used to construct a maximum network flow model. The Dinic maximum flow algorithm is then used to calculate the OCS interconnection topology matching the training traffic demand within the current time slot. Within each OCS scheduling time slot, multiple source nodes and multiple destination nodes corresponding to the aggregated communication are extracted. OCS can connect to link nodes or edges, and super source nodes and super destination nodes are abstractly modeled. In the constructed multi-source, multi-destination network flow graph, the Dinic algorithm is used to calculate the maximum network flow under the current time slot, obtaining the carrying value and priority of each candidate OCS connection under the current training traffic. Based on the traffic allocation results of each source-destination node pair output by the Dinic algorithm, the candidate connections are sorted, prioritizing source-destination node pairs with larger carrying capacity, higher communication urgency, or greater impact on training iteration time to establish OCS optical paths. Training communication tasks are executed under the current OCS topology and routing path, and the remaining data volume, completion status, and subsequent dependent task triggering status of each communication task are updated. When entering the next scheduling time slot, the cross-OCS communication traffic matrix is ​​re-extracted, and step six is ​​repeated until the entire training iteration or complete training process is finished.

[0033] This invention also provides a maximum network flow-based optical circuit switch reconfiguration scheduling training acceleration system, comprising: Workload generation module: Used to generate training workload for large language models based on the input large language model framework, parallel dimension, number of parameters and GPU type parameters; Grouping module: connected to the workload generation module, used to group all GPUs participating in training in the intelligent computing center according to the generated training workload, to obtain data parallel group, tensor parallel group and pipeline parallel group, and to determine the computation and communication tasks within and between groups and the corresponding GPU groups; Directed Acyclic Graph Construction Module: Connected to the grouping module, it is used to create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups. The directed acyclic graph is used to model the order and dependency of computation and communication tasks. Heuristic pruning module: connected to the directed acyclic graph construction module, used to reduce the corresponding parallel dimension and GPU grouping of the directed acyclic graph in the training process with relatively dense cross-optical circuit switch communication traffic through heuristic pruning algorithm, and select the optimal GPU grouping and computation task allocation results; Training execution and traffic monitoring module: connected to the heuristic pruning module, used to execute the complete training computation task and communication task according to the selected optimal GPU group and computation task allocation result, in accordance with the order and dependency relationship of the directed acyclic graph, and generate a traffic matrix of cross-optical circuit switch communication based on the communication of each group set; Optical circuit switch scheduler: connected to the training execution and traffic monitoring module and the optical circuit switch respectively, used to select the optimal optical circuit switch interconnection topology for the corresponding time slot according to the traffic matrix of cross-optical circuit switch communication, and issue the corresponding route, and execute the communication of each group of training tasks under the interconnection topology.

[0034] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0035] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A training acceleration system for optical circuit switches based on maximum network flow reconfiguration scheduling, characterized in that, include: Workload generation module: Used to generate training workload for large language models based on the input large language model framework, parallel dimension, number of parameters and GPU type parameters; Grouping module: connected to the workload generation module, used to group all GPUs participating in training in the intelligent computing center according to the generated training workload, to obtain data parallel group, tensor parallel group and pipeline parallel group, and to determine the computation and communication tasks within and between groups and the corresponding GPU groups; Directed Acyclic Graph Construction Module: Connected to the grouping module, it is used to create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups. The directed acyclic graph is used to model the order and dependency of computation and communication tasks. Heuristic pruning module: connected to the directed acyclic graph construction module, used to reduce the corresponding parallel dimension and GPU grouping of the directed acyclic graph in the training process with relatively dense cross-optical circuit switch communication traffic through heuristic pruning algorithm, and select the optimal GPU grouping and computation task allocation results; Training execution and traffic monitoring module: connected to the heuristic pruning module, used to execute the complete training computation task and communication task according to the selected optimal GPU group and computation task allocation result, in accordance with the order and dependency relationship of the directed acyclic graph, and generate a traffic matrix of cross-optical circuit switch communication based on the communication of each group set; Optical circuit switch scheduler: connected to the training execution and traffic monitoring module and the optical circuit switch respectively, used to select the optimal optical circuit switch interconnection topology for the corresponding time slot according to the traffic matrix of cross-optical circuit switch communication, and issue the corresponding route, and execute the communication of each group of training tasks under the interconnection topology.

2. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The workload includes the overhead of communication categories, communication data volume, and computation time required for each layer of the specific model. The overhead is determined by the type of input GPU, the parallel dimension of model training, the structure of the model, the training framework of the model, the number of parameters of the model, and the parameters of each dimension of the model.

3. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The intra-group and inter-group computation and communication tasks and the corresponding GPU groups are determined according to the framework and parallel dimension of the training process of the large language model. When the execution reaches the corresponding layer, the corresponding computation or communication overhead is written into the workload according to the GPU position of the layer, the input parameters and the parallel dimension. The grouping results are used to determine the mapping relationship between the GPUs participating in each training sub-process in the actual intelligent computing center and the workload.

4. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The directed acyclic graph of the training process is modeled according to the GPU grouping results and the forward and backward propagation of training. Each node of the directed acyclic graph represents the computation or set communication of the corresponding step performed by the corresponding GPU. The edges of the directed acyclic graph represent the dependency relationship between the preceding steps and the subsequent steps. Each node needs to wait for all preceding steps to be completed before it can start execution.

5. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The objective function of the heuristic pruning algorithm is expressed as: in, This represents the communication volume between GPU i and GPU j. Indicates whether cross-OCS communication is required between GPU i and GPU j; The heuristic pruning algorithm avoids cross-optical circuit switch communication by restricting communication-intensive tensor parallel groups to the same rack and reducing cross-domain traffic by allocating data parallel groups to the same optical circuit switch domain. The algorithm sorts candidate schemes according to the heuristic pruning algorithm, removes schemes whose communication cost ranking does not meet the preset requirements, and retains the best or multiple schemes that meet the preset requirements to enter the subsequent training execution and optical circuit switch scheduling optimization process.

6. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The traffic matrix of the cross-optical circuit switch communication originates from the selected communication events of the directed acyclic graph obtained after optimization by a heuristic algorithm. Each communication event includes information on the type of communication, the GPU groups participating in the communication, and the amount of data communicated.

7. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, The maximum network flow algorithm is an improved Dinic maximum network flow algorithm, and its expression is: in, This is the current scheduling slot; It serves as a super source, used to aggregate all sending requests; For a certain source node; To expand the graph edge set, including source edges, destination edges, and candidate optical path edges; Indicates time slot From super source node to source node The allocated traffic.

8. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 7, characterized in that, The process of generating the maximum network flow matching optical circuit switch interconnection topology for the corresponding time slot is as follows: In time slot t, the cross-optical circuit switch communication requirements of the current ready training task are extracted and aggregated into a demand graph. Then, an extended flow graph is constructed by combining the port, plane and bandwidth constraints of the optical circuit switches. Subsequently, the Dinic algorithm is used to solve the maximum flow, and the flow is sorted according to the traffic size of the source node and the destination node pair. High-demand connections are selected first to generate the cost time slot optical circuit switch interconnection topology, and the corresponding routes are issued accordingly.

9. The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow according to claim 1, characterized in that, After entering the next scheduling time slot, the optical circuit switch scheduler re-extracts the traffic matrix of cross-optical circuit switch communication and repeatedly executes the operation of selecting the optimal optical circuit switch interconnection topology for the corresponding time slot through the maximum network flow algorithm until the entire training process is completed.

10. A method for accelerating training reconfiguration scheduling of optical circuit switches based on maximum network flow, characterized in that, The optical circuit switch reconfiguration scheduling training acceleration system based on maximum network flow, as described in any one of claims 1 to 9, comprises the following steps: Step 1: Given the corresponding large language model framework, parallel dimension, parameter size, and GPU type parameters, generate the training workload of the large language model; Step 2: Based on the current training workload, group all GPUs participating in training in the intelligent computing center into data parallel, tensor parallel, and pipeline parallel groups, thereby obtaining different intra-group and inter-group computing and communication tasks and corresponding GPU groups; Step 3: Create a directed acyclic graph of the training process based on the intra-group and inter-group computation and communication tasks and the corresponding GPU groups, which is used to model the sequence and dependencies of computation and communication tasks; Step 4: Use a heuristic pruning algorithm to remove the corresponding parallel dimension and GPU grouping of the directed acyclic graph in the training process where the communication traffic across the optical circuit switch is relatively dense, and select the optimal GPU grouping and computation task allocation results. Step 5: Based on the selected optimal GPU grouping and computation task allocation results, execute the complete training computation task and communication task according to the order and dependency of the directed acyclic graph. In this process, a traffic matrix of cross-optical circuit switch communication based on the communication of each group set is generated. Step 6: Based on the traffic matrix of cross-optical circuit switch communication, generate the maximum network flow matching optical circuit switch interconnection topology for the corresponding time slot according to the improved Dinic maximum network flow algorithm, and execute the communication of each training task under this interconnection topology. After entering the next scheduling slot, repeat step six to complete the entire training process.

Images