1. 1 Performance Improvement of Two-Dimensional Packet Classification by Filter Rephrasing Department of Computer Science and Information Engineering National Cheng Kung University, Taiwan R.O.C. Authors: Pi-Chung Wang, Chun-Liang Lee, Chia-Tai Chan, and Hung-Yi Chang Publisher: IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 15, NO. 4, AUGUST 2007 Present: Yu-Tso Chen Date: February, 12, 2008

2. 2 Outline 1. Introduction 2. Definitions and problem description 3. Tuple Reorganization by Filter Rephrasing 4. Performance Evaluation

3. 3 Introduction Our idea is inspired from the observation that the performance of the tuple-based schemes ties to the length distribution of the prefixes. Hence, we exploit the property of prefix nesting by encoding the prefixes as well as the filters to revise the length distributions.

4. 4 Outline 1. Introduction 2. Definitions and problem description 3. Tuple Reorganization by Filter Rephrasing 4. Performance Evaluation

5. 5 Algorithms for Tuple Space Search

6. 6 Rectangle search algorithm

7. 7 Relationship Between filters and markers The size explosion is caused by an unbalanced distribution of filter lengths

8. 8 Length Distribution of Original Filters

9. 9 Prefix nesting distribution of the routing table The number of nested prefixes for each routing prefix in the existing routing tables from some major network access points

10. 10 Outline 1. Introduction 2. Definitions and problem description 3. Tuple Reorganization by Filter Rephrasing 4. Performance Evaluation

11. 11 Basic Prefix Encoding maximal 7 (=2+3+2) bits are required to represent the original prefixes, but the number of distinct lengths is reduced from 7 to 3.

12. 12 Exclusive Prefix Encoding “Level II” IDs could be reduced to 2 bits and the maximal length is reduced to 6. separate the bitstreams based on their concatenated prefixes at each level

13. 13 Compare the lengths of the encoded prefixes

14. 14 Length Distribution of Encoded Filters The encoded filters only occupy seven tuples The markers are inserted into only four tuples

15. 15 Search Procedure and Implement Heuristic methods Select the field with more distinct prefix lengths as the horizontal axis in the tuple space and perform prefix encoding only on this field. Only one field is encoded, only one BMP lookup is required. Choose the field of destination prefix as the horizontal axis

16. 16 Handling Dynamic Filters Propose a hybrid approach by maintain two tuple spaces 1) For encoded filters 2) For original filters New filters is only inserted into the data structure of the tuple pruning search. Two location lists, one for all filters and one for the new filters. New filter inserted into both location lists.

17. 17 Insertion of Updates Rectangle search and tuple pruning search are performed simultaneously. Once the lookup performance of tuple pruning search degrades to threshold, the new filters are merged into a rect- angle search and trigger reconstruction. During the period, the tuple pruning search looks up all filters by referring to the location list of all filters.

18. 18 Filter Updates

19. 19 Deletion of Updates The pre-computation of the least-cost matching filter is modified to “pre- compute and record all matching filters.” Listed according to their cost in ascending order. Once B is deleted, B and best matching filter A are inserted into the tuple pruning search. And B is recognized as deleted.

20. 20 Example of Filter Deletion

21. 21 Outline 1. Introduction 2. Definitions and problem description 3. Tuple Reorganization by Filter Rephrasing 4. Performance Evaluation

22. 22 Comparison with Rectangle Search Hash table and hash entry present the required storage for 1-D BMP lookups. NLANR routing table

23. 23 Performance Comparisons

24. 24 Filter Length Distribution of synthetic Filter Databases Original filters Encoded filters, marker, and 1-D hash entries Filters and markers

25. 25 Comparison with Rectangle Search(cont.) Synthetic filter databases

26. 26 Performance Comparisons

27. 27 Performance Comparison with Existing Schemes