Abstract:In stateless packet classification, the decision on each packet is independently determined by a rule set, based on its own content. The goal of this dissertation is to improve the cost efficiency, robustness, and performance of stateless packet classification, by exploiting semantic structures of rule sets. In particular, we make the following three contributions. (1) A pure software approach to reducing TCAM storage. Ternary content addressable memory (TCAM) as the de facto solution for wire speed packet classification is expensive and very power consuming. To express rule sets using as little TCAM as possible, previous work proposed to store rules in more compact formats. Thus, in order to transform incoming packets (to match against encoded rules), hardware assistance is needed to avoid possible slowdowns. In contrast, our proposed solution does not change the storage format of rules. Instead, we change the content of rule sets in such a way that will reduce TCAM storage cost but will not alter their classification semantics. Therefore, no additional hardware assistance is needed and no slowdown is incurred. (2) Packet classification without TCAMs. Furthermore, we endeavor to find a solution for wire speed packet classification that does not use TCAMs at all. For that, we propose smart rule cache, which can achieve superior performance using very small cache memory. Such a small cache can be easily implemented in network processors to perform at negligible cost. We believe the value of this solution will only increase as the gap between wire speeds and memory access speeds keeps widening. (3) Approximate packet classification. When overloaded, software-based systems will queue and eventually drop packets. We propose approximate packet classification: if we can quickly approve a considerable percentage of legitimate packets to avoid accumulating packets in the queue, possibly at the cost of mistakenly denying a small percentage of legitimate packets, the overall percentage of legitimate packets that are dropped may still be reduced. Moreover, the average classification delay of approved packets is also much shorter, since they do not have to wait in a long queue.

An Observation of Packet Classification: Most Rules Are at the Top

MiCuts: Combing Bit-Based Cutting and Splitting for Efficient Packet Classification

Packet Classification Algorithms: from Theory to Practice

DOT: Towards Fast Decision Tree Packet Classification by Optimizing Rule Partitions

Performance Evaluation And Improvement Of Algorithmic Approaches For Packet Classification

A fast hash packet classification algorithm

A Multi-Dimensional Packet Classification Algorithm

Towards effective packet classification

CutSplit: A Decision-Tree Combining Cutting and Splitting for Scalable Packet Classification.

Optimized Implementation of Regional Partition Algorithm for Packet Classification

Many Field Packet Classification with Decomposition and Reinforcement Learning

CARR: a Scalable Solution for Network Packet Classification.

An ultra high throughput and memory efficient pipeline architecture for multi-match packet classification without TCAMs.

DBTable: Leveraging Discriminative Bitsets for High-Performance Packet Classification

A Multi-dimensional Packet Classification Algorithm with Trees Divided by Value

PT-Tree: A Cascading Prefix Tuple Tree for Packet Classification in Dynamic Scenarios

A clustered dynamic point split algorithm for packet classification

An Efficient Hybrid Algorithm for Multidimensional Packet Classification.

Poster: CO2: Collaborative Packet Classification for Network Functions with Overselection

DBS: A Bit-level Heuristic Packet Classification Algorithm for High Speed Network

Efficient Schemes for Stateless Packet Classification