1. A Small IP Forwarding Table Using Hashing Yeim-Kuan Chang and Wen-Hsin Cheng Dept. of Computer Science and Information Engineering National Cheng Kung University Tainan, Taiwan R.O.C.

2. Outline Introduction Existing Schemes Designed for Compaction  Techniques used  Performance Evaluation Proposed Data Structure using Hashing  Optimizations  Performance Conclusions

3. Introduction Increasing demand for high bandwidth on Internet  next generation routers: forward multiple millions of packets per second  Required fast routing table lookup

4. Introduction Routing table lookup problem  The most important operation in the critical path  Find the longest prefix that matches the input IP (from more than one matches) Solutions:  Reduces number of memory references  Reduce memory required the forwarding table fit in the application specific integrated circuit (ASIC) fit in the L1/L2 caches

5. Introduction: Binary Trie NGEHOADMICKBFLJ 00010*00010* 0001*0001* 001100*001100* 0100110001001100 0100110*0100110* 0101100101011001 01011*01011* 01*01* 10*10* 1011000110110001 1011001*1011001* 1011001110110011 1011010*1011010* 1011*1011* 110*110* Routing table F is the Longest prefix match for IP = 0101-1000 G E H K ON AB F I M J D C L

6. Existing Schemes Designed for Compaction Small Forwarding Table (Degermark et.al.)  Classified as 12-4-8-8 The compressed 16-x scheme (Huang et.al.)  Classified as 16-x, x = 0 to 16 The Level Compressed Trie (Nilsson et.al.)  Variable stride Technique summary  Run length encoding  Local index of pre-allocation array

7. Memory Sizes # of segments (avg # of prefixes per segment) Sparse: 2,765 (2.9) Dense: 4,300 (25.5) Very dense: 586 (91.7) Maptable: 5.3K Base rray: 2K Code Word array: 8K # of nodes: 259,371 (4 bytes each) Base vector: 110,679 (16 bytes each) Prefix vector: 9,927 (12 bytes each) Next Hop vector: 255 (4 bytes each) Base array: 427 KB Compressed Bit-map: 427 KB CNHA: 37.9 KB # of nodes: 320,478 (7 bytes each) # of nodes: 222,334 (8 bytes each) # of prefixes: 112,286 (4 bytes each) # of blocks: 71,034 (64 bytes each) 5792 nodes of 256 entries (4 bytes each) # of prefixes: 120,635 (45 bits each) Statistics 649.9 KB 2,859 KB 1,147 KB 2,446.8 KB 1,993 KB 1,104 KB 4,695 KB 5,792 KB 662.7 KB Memory size # of level-1 pointers: 13,317 # of level-2 pointers: 461 (4 bytes each) Branch factor: 16 Fill factor: 0.5 16-bit segmentation table: 65535 entries (4 byte each) 8-bit segmentation table N/A Segmentation SFT LC trie Compressed 16-x Binary trie BSD trie Binary range Multiway range Original 8888 Original table Scheme Routing table: 120,635 prefixes.  

8. Lookup latency 1/12 1/5 1/3 8/32 8/26 1/18 1/6 # of memory access (min/max) SFT LC trie Compressed 16-x Binary trie BSD trie Binary range Multiway range Scheme 378 455 230 432 490 264 226 10 th percentile (cycles) 657 777 434 1065 1006 680 541 50 th percentile (cycles) 892 1070 830 1548 1452 1022 861 90 th percentile (cycles) 503 757 409 1140 1077 648 568 Average lookup time (cycles) 0.207 0.312 0.169 0.470 0.444 0.267 0.234 Average lookup time (  s)  

9. Other results for compressed 16-x Memory sizes of compressed 16-x is not proportional to the number of prefixes in the routing table  Number of prefixes with length longer than 24 Therefore, difficult to predict the required amount of memory for compressed 16-x Therefore, we will only compare our proposed scheme with SFT in this paper Oix-80kOix-120kOix-150k 89,088120,635151,511 1,838 Kbytes1,147 KB3,182 Kbytes

10. Proposed Data Structure The binary trie is the simplest data structure Reason for high memory usage in binary trie  space for the left and right pointers.  8 bytes for left and right pointers  1 byte for next port number,  a total of 9 bytes Example routing table: 120,635 entries  ~370K trie nodes in a binary trie. 370K x 9 bytes = 3,330K = 3.3 Mbytes

11. Pre-allocated array of trie nodes If all the trie nodes are organized in an array, we can use node indices of the array as pointers. the trie nodes are physically stored in a sequential array but are logically structured in the binary trie. Assuming no more than one million trie nodes,  20 bits are sufficient for a pointer.  40 bits are required for two pointers plus 8 bits for the next port number.  A total of 6 bytes is required for a trie node. 370K nodes needs around 2.2 Mbytes. Note: LC trie implementation used this idea.

12. Proposed Data Structure (clustering) Simple technique to reduce memory size: Clustering  Nodes in a cluster are layout in an array  Node size can be reduced by using only one global pointer and other pointers are local indices to the array corresponding to the cluster.

13. Proposed Data Structure (clustering) a cluster consists of only one trie node: Node size can be reduced by using only one global pointer. If both left and right children exist, two additional bits are sufficient to indicate whether left and right children exist. The node size is thus reduced from 48 (20*2+8) bits to 30 (20+2+8) bits. (Notice that the number of trie nodes remains the same.) Thus, for the binary trie of 370K trie nodes, the total memory requirement is now 1.4 Mbytes.

14. Proposed Data Structure (clustering) Generalization:  The nodes in a subtree of k levels are grouped into a cluster such that accessing any node inside the cluster is based on the local indices of the node inside the cluster.  Figure I shows the generic data structure for global and local nodes (GNode and LNode). Gnode and Lnode take 31+4k bits and 12+2k bits, respectively. The number N is of size one bit more than M because every node in the cluster can be local node and only the nodes in the bottom of k levels can be global node.

15. Proposed Data Structure (clustering) The number of local nodes increases when k increases.  What is the good choice for k in order to have minimum memory requirement for building the binary trie? The following table shows the memory requirements with different k using a large routing table. We can see that the best value for k is 3. In fact, k = 2 or 4 are also very good. The table also shows the number of internal nodes that are not prefixes, called non-prefix nodes.

16. Memory usage for clustering k levels of the binary trie # of Non-prefix nodes Memory Usage (Mbytes)

17. Multi-bit trie and Hashing Multi-bit trie: improve lookup performance but increasing required memory  Overhead: unused pointers Consider a list of eight 4-bit numbers, 0000, 0001, 0010, 0011, 0100, 1000, 1001, and 1011. A traditional hash function:  H(b n-1 … b 0 ) = ∑b i *2 i  a perfect hash, not minimal 0…00 2 n-1 …21 b i =0 b i =1

18. We can construct arrays  V 0 = [3,0,0,1] and  V 1 = [0,3,2,0]. We have  H(0000) = 4,  H(0001) = 3,  H(0010) = 6,  H(0011) = 5,  H(0100) = 7,  H(1000) = 1,  H(1001) = 0,  H(1011) = 2.

19. A near-minimal perfect hash function The form of the proposed hash function, H, for an n-bit number (b n-1 … b 0 ) is defined as H(b n-1 … b 0 ) = |∑ V b i [i] |, where b i = 0 or 1 and V 0 [n-1 … 0] and V 1 [n-1 … 0] are two precomputed arrays  at least one of V 0 [i] and V 1 [i] is zero and  the values of the non-zero elements are in the range of - MinSize to MinSize,where MinSize = min(H_Size - 1, 2 n-1 ).

20. A near-minimal perfect hash function Computing V 0 [n-1 … 0] and V 1 [n-1 … 0]: The construction algorithm for arrays V 0 and V 1 is based on exhaustive search. Briefly, for each cell in arrays V 0 and V 1, a number in range - MinSize to MinSize is tried one at a time until the hash values of keys are unique.

21. A near-minimal perfect hash function Step 1: sort the keys based on the frequencies of the occurrences of 0 or 1 starting from dimension 0 to n – 1. Now assume the keys are in the order of k 0, …, k m-1 after sorting. In the next two steps the keys are processed in this order.

22. A near-minimal perfect hash function Step 2: compute the cells in arrays V 0 and V 1 that the current key controls. The key, b n-1 … b 0, has the control on V bi [i] if V bi [i] is not yet controlled by one of preceding keys for i = n-1 to 0.

23. A near-minimal perfect hash function Step 3: Use the following rules to assign a number in range (- MinSize ) to MinSize to each cell controlled by the current key.  If the hash value, H(b n-1 … b 0 ), of the key is taken by preceding keys or larger than MinSize or smaller than -MinSize, every cell must be re- assigned a new number.  If no number can be found after exhausting all the possible numbers for the cells controlled by the key, we will backtrack to the previous key by reassigning it a new number and continues the same procedure

24. Analysis of the 4-bit hash table Since the building a hash uses the exhaustive search, it is not feasible for a large n. Therefore, we select the hash table of size n = 4 as the building block for creating a large routing table. We use the exhaustive search to check whether finding a minimal perfect hash function is possible for a set of N 4-bit numbers, where N = 1 to 16.

25. Analysis of the hash table We find out that the minimal perfect hash function exists for all cases except some rare cases when N = 10 or 11. The perfect hash function with the minimal hash size increased by one exists for these rare cases when N = 10 or 11.

26. Example: minimal perfect hash

27. Example: no minimal perfect hash The size of the hash table is one more than that of the minimal prefect hash table. 

28. Lager n The advantage is as follows. The size of the hash table of 8-bit addresses is 10 bytes (8x8+8+8 bits) and thus can fit in a cache block of 16 bytes or larger. However, the obvious drawback is the time- consuming computation of the hash table.  the existence of the minimal perfect hash can not be known in advance,  exponential compute time

29. 4-bit building blocks Additional Data Structure for IP Lookup Store Count and prefix 15(4) 8(9) Numerical representation (1000, 4311)=20bits

30. The 8-8-8-8 Routing table

31. Building and updating Similar to the 4-bit trie except the some internal nodes in each 4-bit subtrie are replaced by the recursive hash tables of 4-bit addresses. Maintain the 4-bit trie and then compute the corresponding hash tables and the associated pointer tables. When a prefix is deleted from or added in the routing table, the 4-bit trie is then updated and thus the corresponding part of the proposed 8-8-8-8 routing table can be changed accordingly.

32. IP lookup performance Based on the distribution of prefix lengths, 99.9% of the routing entries have the prefix length less than or equal to 24. Thus, the number of the memory references is from 1 to 8. The hash tables on level-16 are used to compute the index of level-16 pointer array for accessing the level-24 hashing tables if needed. The operations are similar for level-24 hash tables and pointer table. Notice that we have counted twice of the memory references in accessing the level-16 and level-24 hash tables because accessing two hash tables of 4-bit addresses is needed.

33. Optimizations: Level compression Level compression  The level-8 pointer array can be combined with the left-most hash table of the level-16 hash table, as marked as ‘*a’ in Figure II.  Similarly, the level-16 pointer table can be combined with the left-most hash table of level-24 hash table, as marked as ‘*b’.  Additionally, the level-24 pointer table can be combined with level-32 port table, marked as ‘*c’, but with a possible waste of memory space. With these optimizations, the total number of memory accesses becomes 1 to 5.

34. Optimizations: Maptable Pre-computed hash tables by exhaustive search  217 different hash tables for all possible combinations of N=1..16 4- bit values  (1111, index to the array of pre-computed hash tables): 12 bits (1111, index to array of 16 pre-computed hashed values): avoid computing the hashed values One more memory access to maptable

35. Performance Evaluation Simulation methodology  Input traffic: randomly permutated IP parts of the prefixes in the routing table  Measure the lookup latency one IP at a time  Use the Pentium Instruction rdtsc to get the time in clock cycles

36. Performance Evaluation

38. Average lookup latencies in clock cycles

39. Performance Evaluation Proposed + LC Proposed + LC + Maptable SFT

40. Performance Evaluation Figure 8: IP lookup latency for Oix-120k routing table. Proposed + LC Proposed + LC+maptable Oix-120K

41. Conclusions and future work a new hash table to compact the routing table Hash tables of 4-bit addresses are used as the building blocks to construct 8-8-8-8 hierarchical routing table. Simulations showed  Smaller than any existing schemes  Faster than most schemes except the compressed 16-x scheme Future work: hashing without 4-bit expansion and fast algorithm for n larger than 4

42. The End Thank you