Design and Evaluation of Packet Classification Systems, Doctoral Dissertation, December 2006

Although many algorithms and architectures have been proposed, the design of efficient packet classification systems remains a challenging problem. The diversity of filter specifications, the scale of filter sets, and the throughput requirements of high speed networks all contribute to the difficulty. We need to review the algorithms from a high-level point-of-view in order to advance the study. This level of understanding can lead to significant performance improvements. In this dissertation, we evaluate several existing algorithms and present several new algorithms as well. The previous evaluation results for existing algorithms are not convincing because they have not been done in a consistent way. To resolve this issue, an objective evaluation platform needs to be developed. We implement and evaluate several representative algorithms with uniform criteria. The source code and the evaluation results are both published on a web-site to provide the research community a benchmark for impartial and thorough algorithm evaluations. We propose several new algorithms to deal with the different variations of the packet classification problem. They are: (1) the Shape Shifting Trie algorithm for longest prefix matching, used in IP lookups or as a building block for general packet classification algorithms; (2) the Fast Hash Table lookup algorithm used for exact flow match; (3) the longest prefix matching algorithm using hash tables and tries, used in IP lookups or packet classification algorithms;(4) the 2D coarse-grained tuple-space search algorithm with controlled filter expansion, used for two-dimensional packet classification or as a building block for general packet classification algorithms; (5) the Adaptive Binary Cutting algorithm used for general multi-dimensional packet classification. In addition to the algorithmic solutions, we also consider the TCAM hardware solution. In particular, we address the TCAM filter update problem for general packet classification and provide an efficient algorithm. Building upon the previous work, these algorithms significantly improve the performance of packet classification systems and set a solid foundation for further study