1. EE384Y: Packet Switch Architectures
Part II
Address Lookup and Classification (2)
Nick McKeown
Professor of Electrical Engineering
and Computer Science, Stanford University

2. Outline Routing Lookups Packet Classification
Motivation and problem definition
Classification algorithms
Linear search
Associative search (TCAM)
Trie-based techniques
Crossproducting
Tradeoffs in classification
Heuristic algorithms
References

3. Motivation: Desire for Additional ServicesY
ISP3 X
NAP
ISP1 Z
ISP2
Service
Example
Differentiated Service
Ensure that traffic from ISP2 is given higher priority over traffic from ISP3.
Packet Filtering
Deny all web traffic from ISP3 at interface X.
Policy-based routing
Ensure that all web traffic from ISP2 is sent via interface Z.
Other examples: Accounting & billing, rate-limiting, etc.

4. Special Processing Requires Identification of Flows
All packets of a flow obey a pre-defined rule and are processed similarly by the router
E.g. a flow = (src-IP-address, dst-IP-address), or a flow = (dst-IP-prefix, protocol) etc.
Router needs to identify the flow of every incoming packet and then perform appropriate special processing based on negotiated service agreements
Classification
Rules or policies
(aka ACL entries, filters)

5. Flow-aware Router: Basic Architectural Components
Routing, resource reservation, admission control, SLAs
Control
Packet classification
Special processing
Switching
Datapath:
(per-packet
processing)
Routing lookup
Scheduling

6. Multi-field Packet Classification
L3-DA
L3-SA
L4-PROT
Field 1
Field 2 …
Field k
Action
Rule 1
/21
/32
UDP A1
Rule 2
/24
/16
TCP A2
Rule N
/16 /8
ANY AN
Example: packet ( , , …, TCP)
Packet Classification: Find the action associated with the highest priority rule matching an incoming packet header.

7. Formal Problem Definition
Given a classifier C with N rules, Rj, 1  j  N, where Rj consists of three entities:
A regular expression Rj[i], 1  i  d, on each of the d header fields,
A number, pri(Rj), indicating the priority of the rule in the classifier, and
An action, referred to as action(Rj).
For an incoming packet P with the header considered as a d-tuple of points (P1, P2, …, Pd), the d-dimensional packet classification problem is to find the rule Rm with the highest priority among all the rules Rj matching the d-tuple; i.e., pri(Rm) > pri(Rj),  j  m, 1  j  N, such that Pi matches Rj[i], 1  i  d. We call rule Rm the best matching rule for packet P.

8. Routing Lookup: Instance of 1D Classification
One-dimension (destination address)
Forwarding table  classifier
Routing table entry  rule
Outgoing interface  action
Prefix-length  priority

9. Example 4D Classifier R1 R2 R3 R4 R5 Rule L3-DA L3-SA L4-DP L4-PROT
Action R1 / / *
Deny R2 / /
eq www
udp R3
range 20-21
Permit R4
tcp R5

10. Example Classification Results
Pkt Hdr
L3-DA
L3-SA
L4-DP
L4-PROT
Rule, Action P1
www
tcp
R1, Deny P2
udp
R2, Deny

11. Geometric Interpretation
Packet classification problem: Find the highest priority rectangle containing an incoming point R7 R6 R2 R1 P1 R4 P2 R5 R3
e.g. ( , *)
Dimension 2
e.g. (144.24/24, 64/16)
Dimension 1

12. Outline Routing Lookups Packet Classification
Motivation and problem definition
Classification algorithms
Linear search
Associative search (TCAM)
Trie-based techniques
Crossproducting
Tradeoffs in classification
Heuristic algorithms
References

13. Metrics for Classification Algorithms
Speed
Storage requirements
Ability to handle large classifiers
Low preprocessing time
Update time
Scalability in the number of header fields
Flexibility in rule specification

14. Size/Update-rate of Classifier?
Micro-flow recognition
128K-1M flows in a metro/edge router
Also requires high update rate (but have few wildcards)
Firewall applications
<2K rules per interface
Requires low update rate (usually configured at start-up/boot-up time)
Depends heavily on the type of router

15. Linear Search Keep rules in a linked list
O(N) storage, O(N) lookup time, O(1) update complexity
Q. Why is update complexity O(1) ?

16. Ternary Match Operation
Each TCAM entry stores a value, V, and mask, M
Hence, two bits (Vi and Mi) for each bit position i (i=1..W)
For an incoming packet header, H = {Hi}, the TCAM entry outputs
a match if Hi matches Vi in each bit position for which Mi equals ‘1’. Vi Mi
Match in bit position I ? X
Yes 1
Iff (Hi==0)
Iff (Hi==1)
A simple diagram of a row entry would be useful here.
Z = &_{I=1..W} (~Mi | (Vi == Hi))
Optional Exercise: What is the logic equation for Z (boolean variable denoting whether a TCAM entry matched)?
Optional Exercise: What is the logic equation for Z (boolean variable denoting whether a TCAM entry matched), if instead of (Vi, Mi) we store (Ai,Bi) where (0,0) = always match, (1,1) = always mismatch, (0,1) = match0, and (1,0) = match1

17. Lookups/Classification with Ternary CAM
TCAM
RAM
P32
P31 P8
For LPM
Memory array
Action
Memory
, tcp 1 1 2 3
Priority
Packet
Action
Header
encoder M
1.23.x.x, x 1

18. Range-to-prefix Blowup
Maximum memory blowup = factor of (2W-2)d
Rule
Range R1
[3,11] R2
[2,7] R3
[4,11] R4
[4,7] R5
[1,14]
Maximal Prefixes
0011, 01**, 10**
001*, 01**
01**, 10**
01**
0001, 001*, 01**, 10**, 110*, 1110
Luckily, real-life does not see too many arbitrary ranges.

19. TCAMs Advantages Disadvantages Extensible to multiple fields
Fast: ns today ( M searches per second) going to 250 Msps
Simple to understand and use
Disadvantages
Inflexible: range-to-prefix blowup
Power: 100Msps
Cost: $200-$250 for ~2MByte
Density: largest available in is ~2MB, i.e., 128K x 128 (can be cascaded)
Tough memory soft-error problem

20. Example Classifier Rule Destination Address Source Address R1 0* 10*
01* R3 1* R4
00* R5
11* R6 R7 *

21. Hierarchical Tries Dimension DA O(NW) memory O(W2) lookup Dimension SA
Search (000,010)
Rule DA SA R1 0*
10* R2
01* R3 1* R4
00* R5
11* R6 R7 *
Dimension DA
Dimension SA R5 R2 R1 R3 R6 R7 R4 1
O(NW) memory
O(W2) lookup

22. Set-pruning Tries [Tsuchiya, Sri98]
Search (000,010)
Rule DA SA R1 0*
10* R2
01* R3 1* R4
00* R5
11* R6 R7 *
Dimension DA 1
O(N2) memory
O(2W) lookup R4 R3 R6
Dimension SA R7 R2 R1 R5 R7 R2 R1 R7 R7

23. Grid-of-Tries [Sri98] Dimension DA O(NW) memory O(2W) lookup
Search (000,010)
Rule DA SA R1 0*
10* R2
01* R3 1* R4
00* R5
11* R6 R7 *
Dimension DA 1
O(NW) memory
O(2W) lookup R3 R4 R6
Dimension SA R5 R2 R1 R7

24. Grid-of-Tries
20K 2D rules: 2MB, 9 memory accesses (with prefix-expansion)
Advantages
Good solution for two dimensions
Disadvantages
Difficult to carry out updates
Not easily extensible to more than two dimensions

25. Crossproducting [Sri98]6
(8,4) 5 R2 R1 P1 4 R3 R4
(1,3) 3 2 1 1 2 3 4 5 6 7 8 9

26. Crossproducting Need: d 1-D lookups + 1 memory access, O(Nd) space
50 rules: 1.5MB, need caching (on-demand crossproducting) for bigger classifiers
Advantages
Fast accesses
Suitable for multiple fields
Disadvantages
Large amount of memory
Need caching for bigger classifiers (> 50 rules)

27. Outline Routing Lookups Packet Classification
Motivation and problem definition
Classification algorithms
Linear search
Associative search (TCAM)
Trie-based techniques
Crossproducting
Tradeoffs in classification
Heuristic algorithms
References

28. Classification Algorithms: Speed vs. Storage Tradeoff
Lower bounds for Point Location in N regions with d dimensions from Computational Geometry
O(log N) time with O(Nd) storage, or
O(logd-1N) time with O(N) storage
. These bounds illustrate the difficulty of designing good PC algorithms. Even for small values of N and d, the storage or classification time becomes infeasible.
N = 100, d = 4, Nd = 100 MBytes and logd-1N = 350 memory accesses

29. Classification Tradeoff in Hardware Switches/Routers
Power consumption of classification subsystem
Cost
Speed
Density (Storage)

30. Algorithms so far: Summary
Good for two fields, but do not scale to more than two fields, OR
Good for very small classifiers (< 50 rules) only, OR
Have non-deterministic classification time, OR
Either too slow or consume too much storage

31. One Solution: Heuristics that “seem to work well in real-life”
Recursive Flow Classification [Gupta, McKeown 1999]
Generalization of crossproducting to conserve storage
Hierarchical Intelligent Cuttings [Gupta, McKeown 1999]
Aggregated Bit-vector [Baboescu, Varghese 2001]
Good heuristics do better than worst-case bounds for real-life datasets.
Hierarchy (to at least some level)
Structure
Properties of real-life classifiers:

32. Lookup: What’s Used Out There?
Overwhelming majority of routers:
Modifications of multi-bit tries (h/w optimized trie algorithms)
DRAM (sometimes SRAM) based, large number of routes (>0.25M)
Parallelism required for speed/storage becomes an issue
Others mostly TCAM based
For smaller number of routes (256K)
Used more frequently in L2/L3 switches
Power and cost main bottlenecks

33. Classification: What’s Used Out There?
Majority of hardware platforms: TCAMs
High performance, cost, power, determinstic worst-case
Some others: Modifications of RFC
Low speed, low cost DRAM-based, heuristic
Works well in software platforms
Some others: nothing/linear search/simulated-parallel-search etc.

34. Packet Classification: References
F. Baboescu and G. Varghese, “Scalable packet classification,” Proc. Sigcomm 2001
[Lak98] T.V. Lakshman. D. Stiliadis. “High speed policy based packet forwarding using efficient multi-dimensional range matching”, Sigcomm 1998, pp
[Sri98] V. Srinivasan, S. Suri, G. Varghese and M. Waldvogel. “Fast and scalable layer 4 switching”, Sigcomm 1998, pp [Grid-of-tries, crossproducting]
V. Srinivasan, G. Varghese, S. Suri. “Fast packet classification using tuple space search”, Sigcomm 1999, pp
P. Gupta, N. McKeown, “Packet classification using hierarchical intelligent cuttings,” Hot Interconnects VII, 1999
[Gupta99] P. Gupta, N. McKeown, “Packet classification on multiple fields,” Sigcomm 1999, pp [RFC]

35. Packet Classification: References (contd.)
P. Gupta, “Algorithms for routing lookups and packet classification”, PhD Thesis, Ch 1 and 4, Dec 2000, available at ~pankaj/phd.html [Background and introduction to Classification]
P. Gupta and N. McKeown, “Algorithms for packet classification,” IEEE Network, March/April 2001, vol. 15, no. 2, pp 24-32
S. Iyer, R.R. Kompella, and A. Shelat, “ClassiPI: An architecture for fast and flexible packet classification,” IEEE Network, March/April 2001, vol. 15, no. 2, pp 33-41
TCAM vendors: netlogicmicro.com, sibercore.com, idt.com, cypress.com