ML-Based Fair Flow Control for TCP Core Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing static receive window size (RWND) based bandwidth regulation methods in 5G networks lead to underutilization of resources and inadequate adaptability to dynamic network conditions, resulting in congestion and suboptimal TCP throughput.
Innovation Solution
A machine learning (ML) model is used to predict an optimal window size for user plane gateways based on monitored key performance indicators (KPIs), dynamically adjusting the receive window size to optimize bandwidth allocation and prevent congestion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a static RWND value is used in the UPF, then bandwidth regulation is simplified and easier to implement, but network resource utilization becomes inefficient and adaptability to dynamic network conditions deteriorates
Solution Approach 1:
The patent transforms the static RWND value into a dynamic parameter by introducing an ML-based prediction mechanism. The system continuously monitors network conditions (traffic load, latency, packet loss) and dynamically adjusts the RWND value accordingly, allowing the UPF to adapt to changing network conditions while maintaining automated control.
Solution Approach 2:
The patent implements a feedback loop where the UPF monitors actual network performance metrics (traffic load, latency, packet loss) and uses these measurements to continuously refine the RWND prediction. The ML model learns from historical data and adjusts predictions based on real-time feedback, creating a self-optimizing system that balances simplicity with adaptability.
2Device complexity
If a static RWND value is used in the UPF, then implementation complexity is reduced, but TCP throughput optimization and congestion control effectiveness worsen
Solution Approach 1:
The patent replaces traditional mechanical/manual bandwidth regulation mechanisms with an intelligent ML-based system. Instead of manually configuring static RWND values or using simple threshold-based control, the system employs machine learning algorithms that automatically predict optimal RWND values based on complex network conditions, achieving superior TCP throughput optimization without proportionally increasing implementation complexity.
Solution Approach 2:
The patent enables the UPF to self-optimize TCP throughput by autonomously predicting and adjusting RWND values without requiring manual intervention. The ML model continuously learns from network patterns and automatically adapts to changing conditions, allowing the system to maintain optimal performance while minimizing operational overhead and complexity.
3Speed
If UEs advertise large RWND values due to better hardware capability, then individual UE performance is improved, but network congestion increases and UPF resources become overwhelmed
Solution Approach 1:
The patent applies local quality control by allowing each UE to advertise its hardware-capable RWND value individually, but then applying differentiated regulation at the UPF level. The ML model predicts optimal RWND values tailored to each UE's actual network conditions and capabilities, ensuring that high-capability UEs can utilize their hardware potential while preventing any single UE from overwhelming network resources.
Solution Approach 2:
The patent dynamically changes the RWND parameter based on real-time network conditions and UE-specific factors. Instead of using fixed RWND values or simple caps, the system continuously adjusts the RWND parameter for each UE based on predicted network state, UE hardware capabilities, and current traffic patterns, thereby maintaining reliable congestion control while preserving individual UE performance.
Data Source
AI summary
A method of providing congestion control and reducing latency of data incoming to a core network, the method performed by a control plane gateway, includes: monitoring values of key performance indicators (KPIs) associated with a plurality of user plane gateways in the core network; predicting, using a machine learning (ML) model, an optimal window size respectively for each of the plurality of user plane gateways, based on the monitored values of the KPIs; and transmitting the optimal window size to the respective user plane gateway in the plurality of user plane gateways.