Hash Table Selection Order for Faster Packet Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing packet processing systems in networking devices face inefficiencies due to the need to search multiple hash tables, leading to increased memory access costs and potential collisions, especially with large rule sets, which current strategies like table merging and Bloom filters fail to adequately address.
Innovation Solution
Implementing a machine learning model to dynamically determine the hash table selection order based on packet information, combined with an approximate membership check filter, to optimize the search process and reduce the number of tables accessed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple hash tables are searched to classify packets, then classification accuracy is improved, but processing time and memory access costs increase
Solution Approach 1:
The patent implements dynamic hash table selection where the system adaptively chooses which hash table to search based on packet characteristics and current network conditions. Instead of statically searching all hash tables, the system dynamically determines the optimal search order and selection, reducing unnecessary memory accesses while maintaining classification accuracy.
Solution Approach 2:
The system employs machine learning models that automatically learn and optimize hash table selection strategies without manual intervention. The ML models self-adjust to network traffic patterns and rule set characteristics, enabling the system to serve itself by optimizing its own packet classification performance.
2Measurement precision
If multiple hash tables are searched to classify packets, then classification accuracy is improved, but memory access costs increase
Solution Approach 1:
The system dynamically selects and orders hash tables based on packet characteristics, avoiding unnecessary memory accesses to hash tables that are unlikely to match the current packet. This dynamic approach reduces memory access costs while maintaining the ability to achieve accurate classification when needed.
Solution Approach 2:
The system performs preliminary filtering using approximate membership check filters before searching hash tables. This preliminary action eliminates packets that cannot possibly match any rules, preventing unnecessary memory accesses to hash tables and reducing overall memory access costs.
3Productivity
If hash table search order is optimized using machine learning, then packet classification speed is improved, but device complexity increases
Solution Approach 1:
The patent introduces machine learning models as intermediary components that sit between the packet classification circuits and the hash tables. These ML models act as mediators that analyze packet characteristics and determine optimal hash table search orders, improving classification speed while isolating the complexity of ML algorithms from the core packet processing path.
Solution Approach 2:
The system segments the packet classification function into distinct components: approximate membership check filters, machine learning-based hash table selection, and actual hash table search. This segmentation allows each component to be optimized independently and facilitates implementation using existing hardware structures.
Data Source
AI summary
Network devices, systems, and methods are provided. In one example, a network device includes one or more packet classification circuits and one or more hash selection circuits that cooperate with the one or more packet classification circuits to provide the one or more packet classification circuits with a hash table selection order. The hash table selection order may be determined for a packet to be classified with the support of machine learning by the one or more packet classification circuits based on information contained in the packet.