High-Speed Packet Classification Using Binary

Search on Length Authors: Hyesook Lim and Ju Hyoung Mun Presenter: Yi-Sheng, Lin (林意勝 ) Date: Jan. 14, 2008 Publisher/Conf. : ANCS’07, 2007

Dept. of Computer Science and Information Engineering National Cheng Kung University,

Taiwan R.O.C.

Outline

1. Introduction

2. Area-based quad-trie

3. Binary Search on Prefix Length

4. Proposed Work

5. Optimization Technique

6. Simulation Results

Introduction

We propose an algorithm which applies the binary search on prefix length into the area-based quad-trie for packet classification.

Two new optimization techniques are also proposed.

Area-based quad-trie

Binary Search on Prefix Length

Proposed Work

We propose to separate the area-based quad-trie according to the level of the trie

Storing rules and internal nodes of each level into the corresponding hash table

Performing binary search on those hash tables(Quad-trie table).

Rule table : storing rules with the remaining fields Each entry of the hash table has a rule table pointer whic

h indicates the highest priority rule among the rules mapped into the corresponding node.

Proposed Work

Proposed Work--search(110111,110010,2783,2783,4)

Proposed Work--search

When a node is accessed using the hash key, there could be three cases :

1. Encounter an internal node : guarantees no rule in shorter lengths.

2. Encounter an empty entry (no node) : guarantees no node in longer lengths.

3. Meet a node with rules : Updating best matching rule

and searching can leave the current trie.

Optimization Technique 1

Optimization Technique 2

Simulation Results

The number of rules (N) the number of BSL tries (Nt) The worst-case number of memory accesses (Twst)

The average number of memory accesses(Tavg)

the required memory size in storing BSL tries (Mtrie)

The required memory size in storing a rule table (Mrule) The average memory consumption required in storing a r

ule (M/rule)

Simulation Results

Simulation Results

Simulation Results

Simulation Results

Conclusion

From the simulation result using class-bench databases, we found out that the number of levels of rule nesting in classification tables is 6 at the maximum, and hence the number of tries constructed by the proposed algorithm is limited by 6.

The proposed algorithm showed steady performance not much depending on table characteristics.