# US 2026/0122127 A1 · Distributed Training Method, System, and Apparatus

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0122127 A1
- Document ID: US20260122127A1
- Status: Published application
- Kind: A1
- Application number: 19/003,163
- Filed: 2024-12-27
- Published: 2026-04-30
- Earliest claimed: 2022-06-29
- Assignees: Huawei Cloud Computing Technologies Co., Ltd.
- Applicants: Huawei Cloud Computing Technologies Co., Ltd.
- Inventors: Xiaoyu Zheng; Xibao Pang; Yunwen Lian; Yi Li; Zonghong Dai
- Web page: https://patentgrep.com/patent/US20260122127A1

## CPC

- G06N 3/04: Architecture, e.g. interconnection topology
- H04L 67/1074: for supporting data block transmission mechanisms (file transfer H04L67/06)
- G06N 3/0499: Feedforward networks
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/098: Distributed learning, e.g. federated learning
- G06N 20/00: Machine learning
- H04L 41/0893: Assignment of logical groups to network elements
- H04L 41/12: Discovery or management of network topologies
- H04L 41/16: using machine learning or artificial intelligence
- H04L 45/00: Routing or path finding of packets in data switching networks (routing or path finding in wireless networks H04W40/00)
- H04L 45/02: Topology update or discovery
- H04L 45/12: Shortest path evaluation
- H04L 45/24: Multipath
- H04L 67/1031: Controlling of the operation of servers by a load balancer, e.g. adding or removing servers that serve requests
- H04L 67/1059: {Inter-group management mechanisms, e.g. splitting, merging or interconnection of groups}
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A distributed training method includes a management node that obtains a network topology, where the network topology includes a connection relationship between a core switch and a computing node in a computing cluster. Then, the management node determines a communication plan between N computing nodes based on the network topology, where the N computing nodes are computing nodes that are in the computing cluster and that are configured to train a target model in a distributed manner.

## Claims

### Claim 1

A method comprising:
obtaining a network topology, wherein the network topology comprises a connection relationship between a first core switch and a computing node in a computing cluster, wherein the computing cluster comprises M groups, and wherein each group of the M groups comprises one or more computing nodes;
determining, based on the network topology, a communication plan between N computing nodes that are in the computing cluster and that are configured to train a target model in a distributed manner, wherein the communication plan comprises multiple inter-group paths; and
separately instructing, based on the communication plan, at least two of the N computing nodes to use an inter-group path of the multiple inter-group paths for data communication,
wherein each inter-group path of the multiple inter-group paths comprises:
two computing nodes that are in the N computing nodes and that belong to different groups; and
a second core switch configured to connect the two computing nodes,
wherein the inter-group path is configured to transmit data between the two computing nodes,
wherein an amount of the data meets a preset condition,
wherein M is an integer greater than 2, and
wherein N is an integer greater than 2.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the multiple inter-group paths comprise a first inter-group path, wherein the first inter-group path comprises a first computing node, a second computing node, and a third core switch, wherein separately instructing the at least two of the N computing nodes to use the inter-group path comprises separately sending first information to the first computing node and the second computing node according to the communication plan, and wherein the first information instructs the first computing node to use the first inter-group path to send first data to the second computing node.

### Claim 3 (depends on 1)

The method according to claim 1, wherein each of multiple core switches comprised in the multiple inter-group paths comprises one or more traffic ports, and wherein the preset condition comprises a difference between traffic of any two of multiple traffic ports comprised in the multiple inter-group paths is less than a threshold, and wherein traffic of a traffic port is associated with a data amount of data transmitted between two computing nodes on an inter-group path to which the traffic port belongs.

### Claim 4 (depends on 3)

The method according to claim 3, wherein when each of the multiple inter-group paths comprises multiple levels of core switches, the core switches to which any two traffic ports whose difference is less than the threshold belong to a same level.

### Claim 5 (depends on 1)

The method according to claim 1, wherein for any two inter-group paths of the multiple inter-group paths:
the two inter-group paths comprise different core switches; or
the two inter-group paths comprise a same core switch and traffic ports of the same core switch on the two inter-group paths are different.

### Claim 6 (depends on 1)

The method according to claim 1, wherein the network topology comprises a connection relationship between the first core switch, the computing node in the computing cluster, and an access switch, and wherein for each of the multiple inter-group paths:
the inter-group path further comprises two access switches respectively corresponding to the two computing nodes; and
each computing node on the inter-group path is connected to the second core switch using the access switch corresponding to the computing node.

### Claim 7 (depends on 6)

The method according to claim 6, wherein the communication plan further comprises multiple intra-group paths, and wherein each intra-group path of the multiple intra-group paths comprises:
two computing nodes that are in the N computing nodes and that belong to a same group of the M groups; and
the access switch corresponding to the same group,
wherein the intra-group path is configured to transmit data between the two computing nodes on the intra-group path.

### Claim 8 (depends on 7)

The method according to claim 7, wherein a first data amount of the data transmitted between the two computing nodes on the intra-group path is greater than a second data amount of the data transmitted between the two computing nodes on the inter-group path.

### Claim 9 (depends on 7)

The method according to claim 7, wherein the multiple intra-group paths comprise a first intra-group path, wherein the first intra-group path comprises the first computing node, a third computing node, and a first access switch wherein after determining the communication plan, the method further comprises separately sending second information to the first computing node and the third computing node according to the communication plan, and wherein the second information indicates that the first computing node uses the first intra-group path to send second data to the third computing node.

### Claim 10 (depends on 1)

The method according to claim 1, wherein the M groups respectively correspond to M access switches, and wherein for each access switch of the M access switches:
the access switch comprises K first ports and K second ports respectively corresponding to the K first ports;
the K first ports are respectively connected to K core switches;
the K second ports are respectively connected to K ports of a computing node in a group corresponding to the access switch; and
K is an integer greater than 2.

### Claim 11 (depends on 1)

The method according to claim 1, wherein determining the communication plan comprises determining the communication plan and a communication algorithm, and wherein the communication algorithm is configured to aggregate, in distributed training, data obtained by separately performing training by the N computing nodes in order to obtain the target model.

### Claim 12 (depends on 11)

The method according to claim 11, wherein determining the communication plan comprises:
obtaining a training task comprising a total quantity N of computing nodes and the communication algorithm; and
determining the N computing nodes and the communication plan from multiple computing nodes in an idle state in the computing cluster based on the network topology.

### Claim 13 (depends on 12)

The method according to claim 12, wherein determining the N computing nodes and the communication plan from multiple computing nodes in the idle state in the computing cluster based on the network topology comprises:
determining the N computing nodes from the multiple computing nodes in the idle state in the computing cluster based on the network topology and the total quantity N of computing nodes;
pairing two computing nodes that are in the N computing nodes and that belong to a same group, and when multiple unpaired computing nodes are left, pairing the multiple unpaired computing nodes in order to obtain N/2 node pairs;
determining the communication plan between the N computing nodes in each of multiple rounds of communication based on the multiple rounds of communication of the communication algorithm and the N/2 node pairs, wherein for the communication plan in any round of communication, a larger amount of data transmitted between two computing nodes in the communication plan indicates a smaller quantity of the multiple inter-group paths comprised in the communication plan; and
adjusting, when in an ith round of communication in the multiple rounds of communication, the communication plan comprises the multiple inter-group paths, and an amount of data transmitted on each of the multiple inter-group paths does not meet the preset condition, the communication plan in the ith round of communication, wherein i is a positive integer.

### Claim 14

A system comprising:
a computing cluster comprising M groups, wherein each of the M groups comprises one or more computing nodes;
K core switches configured to connect computing nodes in different groups in the M groups; and
a management node configured to:
obtain a network topology, wherein the network topology comprises a connection relationship between the K core switches and a first computing node in the computing cluster;
determine, based on the network topology, a communication plan between N computing nodes that are in the computing cluster and that are configured to train a target model in a distributed manner, wherein the communication plan comprises multiple inter-group paths; and
separately instruct, based on the communication plan, at least two of the N computing nodes to use an inter-group path of the multiple inter-group paths for data communication,
wherein each inter-path group of the multiple inter-group paths comprises:
two computing nodes that are in the N computing nodes and that belong to different groups; and
a core switch that is in the K core switches and that is configured to connect the two computing nodes,
wherein the inter-group path is configured to transmit data between the two computing nodes,
wherein an amount of the data meets a preset condition, and
wherein K, M, and N are all integers greater than 2.

### Claim 15 (depends on 14)

The system according to claim 14, wherein when determining the communication plan, the management node is further configured to determine the communication plan and a communication algorithm, and wherein the communication algorithm is configured to aggregate, in distributed training, data obtained by separately performing training by the N computing nodes in order to obtain the target model.

### Claim 16 (depends on 14)

The system according to claim 14, wherein each of multiple core switches comprised in the multiple inter-group paths comprises one or more traffic ports, wherein the preset condition comprises a difference between traffic of any two of multiple traffic ports comprised in the multiple inter-group paths is less than a threshold, and wherein traffic of a traffic port is associated with a data amount of data transmitted between two computing nodes on an inter-group path to which the traffic port belongs.

### Claim 17 (depends on 14)

The system according to claim 14, further comprising M access switches respectively corresponding to the M groups, wherein any one of the M access switches is configured to connect a second computing node in a group corresponding to an access switch to the K core switches,
wherein the network topology comprises a connection relationship between the K core switches, the M access switches, and the first computing node in the computing cluster, and
wherein for each of the multiple inter-group paths, the inter-group path further comprises two access switches respectively corresponding to the groups to which the two computing nodes belong.

### Claim 18 (depends on 17)

The system according to claim 17, wherein the communication plan further comprises multiple intra-group paths, and wherein each intra-group path of the multiple intra-group paths comprises:
two computing nodes that are in the N computing nodes and that belong to a same group of the M groups; and
the access switch that is in the M access switches and that corresponds to the group, wherein the intra-group path is configured to transmit data between the two computing nodes on the intra-group path.

### Claim 19 (depends on 14)

The system according to claim 14, wherein the multiple inter-group paths comprise a first inter-group path, wherein the first inter-group path comprises a second computing node, a third computing node, and a first core switch,
wherein the management node is further configured to separately send first information to the second computing node and the third computing node according to the communication plan,
wherein the first information instructs the second computing node to use the first inter-group path to send first data to the third computing node,
wherein the second computing node is configured to send the first data to the first core switch based on the first information,
wherein the first core switch is configured to forward the first data to the third computing node, and
wherein the third computing node is configured to receive the first data from the first core switch based on the first information.

### Claim 20 (depends on 19)

The system according to claim 19, wherein the first inter-group path further comprises a first access switch corresponding to the second computing node and a second access switch corresponding to the third computing node,
wherein the second computing node is further configured to send the first data to the first access switch based on the first information in order to enable the first access switch to send the first data to the first core switch, and
wherein the third computing node is further configured to receive, based on the first information, the first data that is from the first core switch and that is forwarded by the second access switch.
