AI-Aided Flow Identification for Queue-Less Traffic Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing network devices struggle with accurate flow classification and queue-less traffic control, particularly in scenarios where priority queues are absent due to bandwidth differences between local and wide area networks, necessitating innovative solutions for flow identification and traffic control without significant implementation cost increases.
Innovation Solution
A network device employing AI-aided flow identification and drop-based queue-less traffic control, utilizing a flow classification circuit to classify flows into predefined types and a traffic control circuit to apply drop-based control, bypassing the need for priority queues by dynamically adjusting thresholds and drop rates through machine learning and reinforcement learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional flow classification methods are used, then implementation cost is low, but flow classification accuracy is insufficient to distinguish between lower-priority flows and higher-priority flows
Solution Approach 1:
The patent segments the flow classification task into two parts: a lightweight initial classification using traditional methods, and a targeted AI-based refinement for specific flow types. This segmentation allows the system to achieve high classification accuracy only where needed, rather than applying complex AI to all flows, thus improving accuracy while controlling implementation cost.
Solution Approach 2:
The patent introduces an intermediary AI model that acts as a bridge between traditional flow classification and detailed flow type identification. The AI model receives preliminary classified flows and provides override indications only for specific ambiguous cases, serving as an intermediary layer that enhances accuracy without fully replacing the traditional classification system.
2Reliability
If priority queues are used for traffic control, then QoS can be implemented effectively, but the system becomes complex and unsuitable for downstream applications where LAN bandwidth is much larger than WAN bandwidth
Solution Approach 1:
The patent extracts the essential QoS control function from the traditional priority queue system and implements it through a simplified drop-based mechanism. Instead of maintaining complex priority queues and scheduling algorithms, the system extracts the core need for traffic control and realizes it through selective packet dropping based on flow type classification, thereby achieving QoS with reduced complexity.
Solution Approach 2:
The patent changes the control parameter from queue scheduling decisions to packet drop probability. By adjusting the drop rate parameter for different flow types rather than managing multiple priority queues and their scheduling policies, the system achieves traffic control in a simpler manner that is particularly suitable for downstream applications with asymmetric bandwidth characteristics.
3Measurement precision
If AI-based flow identification is applied to all flows, then flow classification accuracy is maximized, but processing overhead and implementation cost increase significantly
Solution Approach 1:
The patent segments the AI processing workload by applying machine learning only to specific flow types that require override indication, rather than analyzing all incoming flows with AI. This selective application of AI to portions of the traffic stream maintains high identification accuracy for critical flows while preserving overall processing efficiency.
Solution Approach 2:
The patent implements partial AI-based flow identification by generating override indications only for specific flows where AI can add value, rather than applying AI universally. This partial action approach achieves sufficient classification accuracy for QoS purposes without the excessive processing overhead that would result from comprehensive AI analysis of all flows.
Data Source
Figure 1
Figure 2~4
Figure 5
AI summary
A network device (100) includes AI aided flow identification circuit (104) and flow classification circuit (106). The AI aided flow identification circuit (104) identifies specific flows of only a portion of pre-defined flow types through machine learning, and generates a flow-type override indication for each of the specific flows. The flow classification circuit (106) classifies each flow into one pre-defined flow type. The AI aided flow identification circuit (104) generates a flow-type override indication for a first flow. In response to the flow-type override indication for the first flow, the flow classification circuit (106) classifies the first flow into a first flow type regardless of a flow rate of the first flow. The AI aided flow identification circuit (104) does not generate a flow-type override indication for a second flow, and the flow classification circuit (106) classifies the second flow into the first flow type or a second flow type according to a flow rate of the second flow.