1. PARALLEL TABLE LOOKUP FOR NEXT GENERATION INTERNET
Author: Parallel Table Lookup for Next Generation Internet
Publisher/Conf.: Computer Software and Applications, COMPSAC '08. 32nd Annual IEEE International
Speaker: Han-Jhen Guo
Date:

2. OUTLINE
Introduction
The Proposed Scheme
Implement
Performance

3. INTRODUCTION - BINARY SEARCH AMONG PREFIX LENGTHS (1/2)
eg. (address length = 8)
Prefix
Nexthop
00* A
1001* B
000* C
001* D
111* E
10010* F 1* G

4. INTRODUCTION - BINARY SEARCH AMONG PREFIX LENGTHS (2/2)
Error!
E (111*) should be matched.
eg. (search )
match
not match
key = 1
key = 11
not match
key = 1110

5. INTRODUCTION - BINARY SEARCH TREE WITH MARKERS (1/3)
Solution: marker
eg.markers of prefix 10010* = 1*, 10*, 100*, 1001*
Meaning: should have a matched prefix longer than this marker
Insert markers into those hash tables in the search path of binary search tree (only pick those markers whose lengths have appeared in the lookup order)
In order to avoid backtracking, the marker is recorded with BMP

6. INTRODUCTION - BINARY SEARCH TREE WITH MARKERS (2/3)

7. INTRODUCTION - BINARY SEARCH TREE WITH MARKERS (3/3)
eg. (search )
match
key = 11
key = 111
not match
match
key = 1110

8. INTRODUCTION - CONCLUSION
The lookup scheme in above is scalable with complexity O(log2W), where W is the length of the IP address.
Assuming that we have a perfect hash function, we only need to do lookup for each hash table only one time
It only need to perform lookup of 5 different hash tables in the worst case in IPv4

9. THE PROPOSED SCHEME - MERGING HASH TABLES
The concept of merging hash tables
(n = 1, 2, 3, 4, etc.)
Assuming either prefix P.0 or prefix P.1 is in Table2n+1 (. means which is followed by a bit) , there should have a marker P in Table2n
Associate a marker P in Table2n with P.0 and P.1
It only need to lookup instead of 4 different hash tables in the worst case after merging

10. THE PROPOSED SCHEME - DATA STRUCTURE OF MODIFIED HASH NODE (1/2)

11. THE PROPOSED SCHEME - DATA STRUCTURE OF MODIFIED HASH NODE (2/2)
eg. after merging

12. THE PROPOSED SCHEME - LOOKUP ALGORITHM
eg. (search )
key = 11
BMP = G → E
key = 1110

13. THE PROPOSED SCHEME - MAKING LOOKUP ALGORITHM PIPELINED (1/2)
Binary search tree for IPv6 without merging hash tables
Modified binary search tree for IPv6

14. THE PROPOSED SCHEME - MAKING LOOKUP ALGORITHM PIPELINED (2/2)
assign each one to do lookup of the hash table in one level → 6 stages totally
Each stages:
use the destination IP address as the key to do hash
lookup the hash table using the hash value as the index
do computation according to the lookup result
BMP so far
the hash table to be searched next
the skip flag for the next processing unit regarding the BMP that has been found

15. THE PROPOSED SCHEME - USING MULTI-THREADING IN THE PIPELINE STAGE

16. IMPLEMENT - IMPLEMENTATION PLATFORM
IXP2400
8 micro engines; it supports 8 threads each micro engine
use 6 micro engines to implement our design of pipeline, and run 8 threads on each micro engine for realizing the design of multi-threading
IXA SDK 4.1
to simulate the environment of IXP2400

17. IMPLEMENT - IMPLEMENTATION BRIEFS (1/4)
Maximum size of three separate memories
The average latencies of reading eight 4-byte words from SRAM and DRAM in the circumstance of only one micro engine trying to access the memories

18. IMPLEMENT - IMPLEMENTATION BRIEFS (2/4)
The average latencies of reading 8 words from a certain channel of SRAM or DRAM when different numbers of micro engines try to contend for accessing that channel of SRAM or DRAM
allow 8 simultaneous SRAM accesses (4 from each channel) and 3 simultaneous DRAM accesses without increasing the average memory latency

19. IMPLEMENT - IMPLEMENTATION BRIEFS (3/4)
Distribute hash tables to 3 separate memory of IXP2400

20. IMPLEMENT - IMPLEMENTATION BRIEFS (4/4)
Hashing
hash function: CRC32
collision resolution: chaining
alleviate the penalty of hash collision with access 2 contiguous nodes a time with a little bit of memory latency
Fig. Chaining in a hash
Fig. Chaining with 2 contiguous nodes

21. PERFORMANCE Comparisons of maximum forwarding rates
10,000 random IP addresses and calculate the number of total cycle counts required to perform lookup

22. Thanks for
your listening!