Removing Redundancy from Packet Classifiers

Packet classification is the core mechanism that enables many networking services such as firewall access control and traffic accounting. Reducing memory space for packet classification algorithms is of paramount importance because a packet classifier must use very limited on-chip cache to store complex data structures. This paper proposes the first ever scheme that can significantly reduce memory space for all packet classification algorithms. The scheme is to remove all redundant rules in a packet classifier before a classification algorithm starts building data structures. By removing redundant rules, we can save more than 73% of memory for a packet classifier that examines eight packet fields. In this paper, we categorize redundant rules into upward redundant rules and downward redundant rules. We give a necessary and sufficient condition for identifying each type of redundant rule. We present two efficient algorithms for detecting and removing the two types of redundant rules respectively. The two algorithms make use of a graph model of packet classifiers, called packet decision diagrams. The experimental results shows that our algorithms are very efficient.