Layered Cutting Scheme for Packet Classification

Abstract

Packet classification is an important topic for high speed routers nowadays. There are many packet classification algorithms based on decision tree like Hicuts, Hyper cuts and Hyper split. Because Hicuts and Hyper cuts divides the rule sets by cutting the address space into equal-sized subspaces, their cutting efficiency is not good. Although Hyper split proposed a good end-point-based cutting scheme, the resulting tree depth is still very high. In this paper, we propose a multi-dimensional cutting algorithm to significantly reduce the decision tree depth and a multi-layered scheme to dramatically reduce the usage of memory. Our experimental results show that the proposed layered scheme needs much less memory than Hyper split for Firewall and IPC rule tables with a factor of 2 to 106 improvement while the proposed layered scheme needs a little more memory than Hyper split for some of ACL tables. In addition, in terms of number of memory accesses, the proposed layered scheme and Hyper cuts are better than Hicuts and Hyper split for all tables while the proposed layered scheme is better than Hyper cuts for ACL and Firewall tables. In terms of number of memory accesses, our layered cutting scheme and Hyper cuts perform equally well for small rule tables. But, in larger rule tables, the proposed layered cutting scheme has better performance.