Flow Scheduling Using Spine Port Cycle Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI large model networks experience traffic congestion due to periodic fluctuations and large data volumes, leading to traffic collisions at Spine nodes, particularly when multiple computing nodes send traffic to the same Leaf node.
Innovation Solution
A traffic forwarding method and apparatus that involves a forwarding engine reporting flow statistics information to a processor, merging flow statistics based on timestamps, determining cycle information, and advertising flow characteristic information to Leaf nodes for upstream data flow scheduling, thereby reducing congestion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple computing nodes simultaneously send traffic to computing nodes under the same Leaf node, then data transmission volume increases, but traffic collisions and congestion occur at the downstream port of the Spine node
Solution Approach 1:
The system performs preliminary actions by collecting flow statistics information in advance and determining cycle information before traffic congestion occurs. The processor determines cycle duration, traffic interval duration, and traffic rate ahead of time, enabling proactive scheduling decisions that prevent congestion rather than react to it.
Solution Approach 2:
The system implements feedback mechanisms by advertising flow characteristic information back to leaf nodes based on Spine node observations. The flow characteristic information includes cycle information and traffic patterns, which leaf nodes use to adjust their upstream data flow scheduling, creating a closed-loop control system that continuously optimizes traffic distribution.
2Reliability
If flow statistics information is collected and processed in real-time, then traffic congestion can be reduced, but network overhead and processing complexity increase
Solution Approach 1:
The system merges flow statistics information with similar characteristics into unified cycle information representations. Instead of processing each individual flow statistic separately, the processor aggregates them into cycle duration, traffic interval duration, and traffic rate parameters, significantly reducing processing complexity while maintaining congestion reduction effectiveness.
Solution Approach 2:
The system transforms detailed flow statistics into simplified cycle information parameters. By changing the representation from raw packet-level statistics to aggregated temporal and volumetric parameters (cycle duration, traffic interval, traffic rate), the system reduces processing complexity while preserving the essential information needed for congestion prevention.
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
Disclosed are a traffic forwarding method and apparatus. In an example of the present disclosure, a forwarding engine may report flow statistics information to a processor in response to a determination that an information reporting condition is met. Among flow statistics information with the same flow identification information, the processor may merge the flow statistics information meeting a preset interval duration merging condition based on the first packet timestamp and the last packet timestamp. The processor determines cycle information of a flow based on the merged flow statistics information, determine flow characteristic information of a downstream port of a spine node based on the cycle information of the flow, and advertise the flow characteristic information of the downstream port of the spine node to leaf nodes, so that the leaf nodes perform upstream data flow scheduling based on the flow characteristic information of the downstream port of the spine node.