Virtual NIC Routing for Cross-Subnet AI Training Communication

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In AI training platforms, ROCE network cards in different sub-networks cannot communicate directly, leading to potential failure of training tasks due to broadcast storms and the need for unified network management, especially when using containers.

Innovation Solution

A network card communication method that builds a spine-leaf network topology, virtualizes physical network cards into virtual ones, allocates them to job-containers with specific sub-networks, and adds sub-network communication policy routing rules to ensure data transmission across different sub-networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If multiple ROCE network cards are divided into different sub-networks to avoid broadcast storms and comply with unified network management, then network management is improved, but cross sub-network communication fails

Engineering Contradiction:
Improvenetwork managementVSAvoidcross sub-network communication
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent introduces a routing rule mechanism as an intermediary that enables communication between different sub-networks. The routing rule includes a source sub-network identifier and destination sub-network identifier, allowing the system to route data packets between sub-networks through a mediator mechanism, thus resolving the communication failure issue while maintaining the benefits of sub-network division

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the network into multiple sub-networks for better management, and then introduces a routing mechanism that segments the communication path between sub-networks. Each sub-network maintains its independence for management purposes while the routing rules create controlled communication paths between them, resolving the contradiction between segmentation benefits and communication capability

Inventive Principle:
Principle #1Segmentation

2Object-affected harmful factors

If NCCL_CROSS_NIC=0 is set to avoid cross sub-network communication, then broadcast storms are avoided, but training task reliability deteriorates

Engineering Contradiction:
Improvebroadcast stormsVSAvoidtraining task
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The routing rule acts as an intermediary mechanism that selectively enables cross sub-network communication only when needed for training tasks, while maintaining the default isolation that prevents broadcast storms. The routing rule includes source and destination sub-network identifiers that control when communication is permitted, thus preventing harmful broadcast storms while enabling reliable training task completion

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If physical network cards are virtualized and allocated to different sub-networks, then resource utilization is improved, but communication reliability deteriorates

Engineering Contradiction:
Improveresource utilizationVSAvoidcommunication
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The routing rule mechanism serves as an intermediary that ensures reliable communication between virtual network cards allocated to different sub-networks. The routing rule includes source sub-network identifier and destination sub-network identifier, enabling the system to route data packets correctly between different sub-networks, thus maintaining communication reliability while achieving improved resource utilization through virtualization

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentUS12500834B2Network card communication method and apparatus for AI training platform, device, and medium
Publication Date: 2025.12.16 INSPUR SUZHOU INTELLIGENT TECH CO LTD
  • US12500834B2 patent drawing
  • US12500834B2 patent drawing
  • US12500834B2 patent drawing

AI summary

Disclosed are a network card communication method and apparatus for an AI training platform, a device, and a medium. The method includes: building a switch network topology on the basis of a spine-leaf network, and configuring a preset number of virtual local area networks for each leaf switch in the switch network topology; virtualizing a physical network card to obtain virtual network cards, allocating the virtual network cards to corresponding job-containers according to a preset allocation rule, and allocating, to each virtual network card in the job-containers, different sub-networks corresponding to the virtual local area networks; and adding a corresponding sub-network communication policy routing rule to a pod where each job-container is located, whereby a virtual network card in the job-container sends training data to the remaining virtual network cards on a basis of the sub-network communication policy routing rule.