Class-Based Bandwidth Partitioning via Probabilistic Dropping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current bandwidth partitioning methods, such as Distributed Round Robin (DRR) and Weighted Random Early Detection (WRED), face challenges in scalability and fairness, with Approximate Fair Dropping (AFD) providing theoretical but not comprehensive practical solutions for class-based bandwidth allocation in packet switching systems.
Innovation Solution
A method for class-based bandwidth partitioning that determines packet admission or dropping based on probabilities associated with measured arrival rates and fair shares calculated from queue lengths, periodically updating these parameters to dynamically adjust to changing traffic conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If Distributed Round Robin (DRR) is used for bandwidth partitioning, then quantum-level fairness is achieved, but device complexity and scalability are worsened due to the need to store traffic in separate queues and implement complex round-robin scheduling
Solution Approach 1:
The patent segments the single FIFO buffer into multiple virtual queues through software classification, where each packet is tagged with a class identifier. This allows DRR-like fairness to be achieved without maintaining separate physical queues for each traffic class, reducing memory overhead and complexity while preserving quantum-level fairness through probabilistic dropping based on class-specific parameters.
Solution Approach 2:
The patent makes the single FIFO buffer serve multiple functions by implementing class-based probabilistic dropping within it. The buffer simultaneously handles multiple traffic classes and provides fairness guarantees for all classes without requiring separate physical queue structures, thus achieving multi-functionality with a single data structure.
2Ease of operation
If Weighted Random Early Detection (WRED) is used for bandwidth partitioning, then ease of operation is improved due to simple single FIFO buffer implementation, but bandwidth fairness and adaptability are worsened due to limited fairness and non-self-adjusting parameters
Solution Approach 1:
The patent introduces dynamic parameters into WRED by making the dropping probability class-dependent and periodically updated based on measured arrival rates and queue lengths. Each traffic class has its own dropping probability parameter that adapts to changing traffic conditions, transforming static WRED into a dynamic system that maintains fairness while keeping the simple single buffer structure.
Solution Approach 2:
The patent implements feedback mechanisms where the system periodically measures packet arrival rates for each class and current queue lengths, then uses this information to update the dropping probability parameters. This closed-loop feedback allows the system to self-adjust and maintain bandwidth fairness without requiring complex manual parameter tuning, bridging the gap between WRED's simplicity and DRR's fairness.
3Manufacturing precision
If Approximate Fair Dropping (AFD) is used for bandwidth partitioning, then bandwidth fairness is improved with limited additional states, but comprehensive practical implementation is worsened due to lack of complete solution for class-based allocation
Solution Approach 1:
The patent applies local quality by implementing class-specific dropping probabilities within the single buffer framework. Each traffic class can have different dropping parameters tailored to its characteristics, allowing differentiated treatment of different packet classes while maintaining the simplicity of a single physical buffer. This enables comprehensive class-based allocation with fine-grained control.
Data Source
Figure 1
Figure 2~3A
Figure 3B
AI summary
Class-based bandwidth partitioning of a sequence of packets of varying packet classes is performed, such as, but not limited to determining whether or not to admit a packet to a queue based on a probability corresponding to a class of packets associated with the packet, with this probability being based on measured arrival traffic and a fair share based on the length of the queue. Data path processing is performed on each packet to determine whether to admit or drop the packet, and to record the measured received traffic. Control path processing is periodically performed to update these probabilities based on determined arrival rates and fair shares for each class of packets. In this manner, a relatively small amount of processing and resources are required to partition bandwidth for a scalable number of classes of packets.