1. An Efficient Hardware-based Multi-hash Scheme for High Speed IP Lookup Department of Computer Science and Information Engineering National Cheng Kung University, Taiwan R.O.C. Authors: Socrates Demetriades, Michel Hanna, Sangyeun Cho and Rami Melhem Publisher: High Performance Interconnects, 2008. HOTI '08. 16th IEEE Symposium on 26-28 Aug. 2008 Present: Chia-Ming,Chuang Date: 9, 24, 2008

2. Outline 1. Introduction 2. Architecture 3. Bit select mechanism 4. Conclusion

3. Introduction (1/3) Ternary Content Addressable Memory (TCAM) Trie-based schemes Hash-based schemes Current techniques IP lookup have three categories

4. Introduction (2/3) First hashing can happen collisions and requires resolution techniques that may result in unpredictable lookup performance Second hash keys be included “don’t care” bit Three determine the longest prefix hash-based IP lookup problem

5. Introduction (3/3) High lookup throughput A multiple table, multiple hash functions Bit select mechanism high space utilization Modified version of the Cuckoo hashing algorithm Low power consumption Sum of advantage

6. Outline 1. Introduction 2. Architecture 3. Bit select mechanism 4. Conclusion

7. Architecture overview (1/4)

8. Architecture (2/4) Based on a multi-level hash table (MHT) All hash function use the same bit position in a prefix or a destination address PS:references [16] S. Cho, J. Martin, R. Xu, M. Hammoud, and R. Melhem, “Ca-ram: A high-performance memory substrate for search-intensive applications,” pp. 230–241, IEEE ISPASS, April 2007.

9. Architecture (3/4)

10. Architecture (4/4) Do not classify prefixes based on their lengths Each table is accessed using a hardware based index generator a number of match processors compare the fetched keys with the search key in parallel “Bit-Select” component (1)decides which bits participate in the index generation (2)“don’t care bits” may appear in the selected bits for indexing (3)Bit-Select mechanism needs to resolve this by prefix expansion

11. Outline 1. Introduction 2. Architecture 3. Bit select mechanism 4. Conclusion

12. Bit select mechanism (1/7)

13. Bit select mechanism (2/7) ( 一 ) Most prefix have lenghts between16 and 24 bits 24 bit prefixes comprise about 54% no prefixes less than 8 bit ( 二 ) As a study in [9] pointed out, regardless of prefix length, start from the6th bit and reach the prefixes’ maximum length.

14. Bit select mechanism (3/7)

15. 15 Bit select mechanism (4/7) ( 一 ) X-axis =>(Expansion size)/extra provided capacity Y-axis=> unresolved collisions no prefixes less than 8 bit ( 二 ) From the experiments, we find that any configuration that can keep the expansion between 20% and 90% of the extra provided capacity results in less than 5% of unresolved collisions.

16. Bit select mechanism (5/7) controlled prefix expansion controlled wildcard resolution

17. 17 Max bucket is 234 item Max bucket is 49 item Total 49*3=147 Cuckoo hashing algorithm Bit select mechanism (6/7)

18. 18 Bit select mechanism (7/7) ( 一 ) a collision may appear if an item x can not be inserted into tables Ti, i = 1,..., d because buckets Ti [hi(x)], i = 1,..., d are full. ( 二 ) by checking whether an item y at some bucket Ti [hi(x)], 1 ≤ i ≤ d, can migrate to some other table Tj =\ Ti. This requires that a bucket Tj [hj(y)] has an empty entry to accommodate the item y. ( 三 ) If such an item y is found, then it can be moved and x will take its place.

19. Outline 1. Introduction 2. Architecture 3. Bit select mechanism 4. Conclusion

20. Conclusion (1/3)

21. Conclusion (2/3)

22. Conclusion (3/3)