Packet classification using coarsegrained tuple spaces

1
Packet classification using coarse-grained tuple
spaces
Authors Haoyu Song, Jonathan Turner, Sarang
Dharmapurikar Publisher Proceedings of the 2006
ACM/IEEE symposium on Architecture for networking
and communications systems Present Min-Yuan
Tsai (???) Date January, 18, 2007
Department of Computer Science and Information
Engineering National Cheng Kung University,
Taiwan R.O.C.
2
Outline

• 1. Introduction
• 2. Combining The Two Algorithms
• 3. Longest Prefix Matching
• 4. Evaluation
• 5. Conclusion

3
Introduction

• Some disadvantages of current solutions for
  packet classification
• The fast algorithmic methods require excessive
  space
• TCAM has high cost per bit, significant power
  usage
• The method of this paper is combining
  cross-product and tuple-space search
• Improving one-dimensional(1d) lookups
• combining tree bitmap and Bloom filters

4
Outline

• 1. Introduction
• 2. Combining The Two Algorithms
• 3. Longest Prefix Matching
• 4. Evaluation
• 5. Conclusion

5
Combining The Two Algorithms

• Cross-Product Method Procedure
• do 1d lookup on all fields
• combine results into lookup key in cross-product
  table
• cross-product table direct lookup table or hash
  table

10100, 01110
S0D1
6
Combining The Two Algorithms (contd.)

• Cross-Product Method Expansion
• Assume the prefix pair (A,B) is an original
  filter, R1. If the prefix pairs (A,D), (C,B), or
  (C,D) are not original filters, we need to add
  pseudo-filters for each to guarantee correct
  lookups.
• For example, if the single
• field lookups return the
• matching prefixes C and
• D, the results imply a
• match to the filter R1.
• Without the pseudo-filter
• (C,D), we will miss this
• match.

7
Combining The Two Algorithms (contd.)

• Cross-Product Method performance
• We can also see from the last column of Table 2
  that 32 to 87 of entries in the direct lookup
  table actually do not contain any matching
  filter.
• This fact suggests that we use a hash table
  instead of a direct lookup table for better
  storage efficiency.

8
Combining The Two Algorithms (contd.)

• 2D Tuple Space Search
• Group by prefix length
• hash table per group
• up to 33 x 33 1,089 groups
• Rectangle search
• markers to guide search
• at most 63 probes, often less
• Pruned tuple space search
• 1d search on src/dest fields
• find prefix lengths that match src/dest fields of
  packet
• search intersecting tuples
• if w matching prefixes, at most w2 probes

9
Combining The Two Algorithms (contd.)

• Coarse-Grained Tuple Space
• Select coarse-grained partition of tuple space
• Build cross-product table per sub-space
• Search procedure
• 1d lookups for LPM
• identify candidate sub-spaces during 1d lookup
• probe selected sub-spaces
• terminate early if possible

10
Combining The Two Algorithms (contd.)

• Coarse-Grained Tuple Space Example
• A given packet
• source IP matching prefix in segments 0, 7 and
  17, 32, destination IP matching prefix in
  segments 9, 21 and 22, 32.
• The best matching filter can
• exist only in the hash tables for
• the tuple sets D, F, G, and I.
• Now the logical choice is to start the search
  from the tuple sets I,
• G, and F, because if we find matching filters
  belonging to
• these tuple sets, we do not need to search the
  tuple set D.

17
7
A
B
C
8
D
E
F
21
G
H
I
11
Outline

• 1. Introduction
• 2. Combining The Two Algorithms
• 3. Longest Prefix Matching
• 4. Evaluation
• 5. Conclusion

12
Longest Prefix Matching
Tree Bitmap
Hashing Bloom Filters

• Tree Bitmap is a multibit trie algorithm that
  allows fast searches
• Bitmaps for
• prefix nodes
• subtree presence

• Expand prefixes to standard lengths
• Off-chip hash table per length
• The lookup process probes all the Bloom filters
  simultaneously and uses the output of the Bloom
  filters to determine which hash table to query.

13
Longest Prefix Matching (contd.)
Bloomfilters
Each encoded subtree is called a Segment Multibit
Trie (SMT)

• Instead of performing the prefix expansion, this
  paper encode the subtree between two length
  thresholds using the Tree BitMap or Shape
  Shifting Tries encoding technique. (a multi-bit
  trie algorithm with multilevel jump tables)
• Insert tree bitmap subtree roots into off-chip
  hash tables
• Lookup prefix of subtree roots in Bloom filters
• if match on length k and all shorter lengths,
  probe off-chip table for length k

14
Longest Prefix Matching (contd.)

• 200K prefixes in the route table

5 bit stride for tree bitmap Using only 5 Bloom
filters
Reducing the number of Bloom filters
15
Outline

• 1. Introduction
• 2. Combining The Two Algorithms
• 3. Longest Prefix Matching
• 4. Evaluation
• 5. Conclusion

16
Evaluation
a ß means that the source IP address field has
a equal sized segments and the destination IP
address field has ß equal sized segments. a ß
is the worst-case number of hash queries for a
packet
Ratio of cross-products to filter set size

• Equal size divisions of 2d tuple space
• 2x2 partition brings space usage to 2x minimum
• maximum of four probes required
• For 2d filter sets, the best-matching filter with
  just 4 hash probes while limiting the memory
  space expansion factor to about 2.

17
Evaluation (contd.)
Original size of the filter size
18
Outline

• 1. Introduction
• 2. Combining The Two Algorithms
• 3. Longest Prefix Matching
• 4. Evaluation
• 5. Conclusion

19
Possible Extensions

• This paper present a novel packet classification
  algorithm derived from the tuple space search
  algorithm and the cross-producting algorithm.
• Future work
• Generalization to more dimensions
• handling fields with ranges (for port numbers)
• Tuple partition
• The simple tuple partition seems to work well on
  current filters, it make sense to study better
  tuple partition algorithm for larger filter sets
  in the future
• Is this method better than TCAM?