1

Algorithms for Packet

Classification

Pankaj Gupta and Nick McKeown

Computer Systems Laboratory, Stanford University
Stanford, CA 94305-9030
{pankaj, nickm}@stanford.edu

Abstract

The process of categorizing packets into “flows” in an Internet router is called packet
classification. All packets belonging to the same flow obey a pre-defined rule and are processed in
a similar manner by the router. For example, all packets with the same source and destination IP
addresses may be defined to form a flow. Packet classification is needed for non “best-effort”
services, such as firewalls and quality of service; services that require the capability to distinguish
and isolate traffic in different flows for suitable processing. In general, packet classification on
multiple fields is a difficult problem. Hence, researchers have proposed a variety of algorithms
which, broadly speaking, can be categorized as “basic search algorithms,” geometric algorithms,
heuristic algorithms, or hardware-specific search algorithms. In this tutorial we describe
algorithms that are representative of each category, and discuss which type of algorithm might be
suitable for different applications.

1 Introduction
Until recently, Internet routers provided only “best-effort” service, servicing packets in
a first-come-first-served manner. Routers are now called upon to provide different quali-
ties of service to different applications which means routers need new mechanisms such as
admission control, resource reservation, per-flow queueing, and fair scheduling. All of
these mechanisms require the router to distinguish packets belonging to different flows.

Flows are specified by rules applied to incoming packets. We call a collection of rules
a classifier. Each rule specifies a flow that a packet may belong to based on some criteria
 2

PAYLOAD L4-SP L4-DP L4-PROT L3-SA L3-DA L3-PROT L2-SA L2- DA

16b 16b 8b 32b 32b 8b 48b 48b

Transport layer header Network layer header Link layer header

L2 = Layer 2 (e.g., Ethernet) DA =Destination Address

L3 = Layer 3(e.g., IP) SA = Source Address
L4 = Layer 4(e.g., TCP) PROT = Protocol
SP = Source Port
DP =Destination Port
Figure 1 This figure shows some of the header fields (and their widths) that might be used for classifying
the packet. Although not shown in this figure, higher layer (e.g., application-level) headers may also be
used.

applied to the packet header, as shown in Figure 1. To illustrate the variety of classifiers,
consider some examples of how packet classification can be used by an ISP to provide dif-
ferent services. Figure 2 shows ISP1 connected to three different sites: enterprise networks
E1 and E2 and a Network Access Point1 (NAP), which is in turn connected to ISP2 and
ISP3. ISP1 provides a number of different services to its customers, as shown in Table 1.
TABLE 1.

Service Example
Packet Filtering Deny all traffic from ISP3 (on interface X) destined to E2.

Policy Routing Send all voice-over-IP traffic arriving from E1 (on interface Y) and
destined to E2 via a separate ATM network.

Accounting & Billing Treat all video traffic to E1 (via interface Y) as highest priority and
perform accounting for the traffic sent this way.

Traffic Rate Limiting Ensure that ISP2 does not inject more than 10Mbps of email traffic
and 50Mbps of total traffic on interface X.

Traffic Shaping Ensure that no more than 50Mbps of web traffic is injected into ISP2
on interface X.

1. A network access point is a network site which acts as an exchange point for Internet traffic. ISPs connect to the NAP
to exchange traffic with other ISPs.
 3

E1
Y

ISP1

ISP2 X
Z
NAP
E2

ISP3 Router

Figure 2 Example network of an ISP (ISP1) connected to two enterprise networks (E1 and E2) and to two
other ISP networks across a network access point (NAP).

Table 2 shows the flows that an incoming packet must be classified into by the router
at interface X. Note that the flows specified may or may not be mutually exclusive. For
TABLE 2.

Flow Relevant Packet Fields:

Email and from ISP2 Source Link-layer Address, Source Transport port number

From ISP2 Source Link-layer Address

From ISP3 and going to E2 Source Link-layer Address,

Destination Network-Layer Address
All other packets —

example, the first and second flow in Table 2 overlap. This is common in practice, and
when no explicit priorities are specified, we follow the convention that rules closer to the
top of the list take priority.

1.1 Problem statement

th
Each rule of a classifier has d components. R [ i ] is the i component of rule R, and is
th
a regular expression on the i field of the packet header. A packet P is said to match rule
th
R, if ∀i , the i field of the header of P satisfies the regular expression R [ i ] . In practice, a
 4

rule component is not a general regular expression but is often limited by syntax to a sim-
ple address/mask or operator/number(s) specification. In an address/mask specification, a
0 (respectively 1) at bit position x in the mask denotes that the corresponding bit in the
address is a don’t care (respectively significant) bit. Examples of operator/number(s) spec-
ifications are eq 1232 and range 34-9339. Note that a prefix can be specified as an address/
mask pair where the mask is contiguous — i.e., all bits with value 1 appear to the left of
t
bits with value 0 in the mask. It can also be specified as a range of width equal to 2 where
t = 32 – prefixlength . Most commonly occurring specifications can be represented by

ranges.

An example real-life classifier in four dimensions is shown in Table 3. By convention,

the first rule R1 is of highest priority and rule R7 is of lowest priority. Some example clas-
sification results are shown in Table 4.
TABLE 3.

Network-layer Transport- Transport-

Network-layer Source
Rule Destination (address/ layer layer Action
(address/mask)
mask) Destination Protocol

R1 152.163.190.69/ 152.163.80.11/ * * Deny

255.255.255.255 255.255.255.255

R2 152.168.3.0/ 152.163.200.157/ eq www udp Deny

255.255.255.0 255.255.255.255

R5 152.163.198.4/ 152.163.160.0/ gt 1023 tcp Permit

255.255.255.255 255.255.252.0

R6 0.0.0.0/0.0.0.0 0.0.0.0/0.0.0.0 * * Permit

TABLE 4.

Transport- Transport-
Packet Network-layer Network-layer Best matching
layer layer
Header Destination Source rule, Action
Destination Protocol

P1 152.163.190.69 152.163.80.11 www tcp R1, Deny

P2 152.168.3.21 152.163.200.157 www udp R2, Deny

P3 152.168.198.4 152.163.160.10 1024 tcp R5, Permit

 5

Longest prefix matching for routing lookups is a special-case of one-dimensional

packet classification. All packets destined to the set of addresses described by a common
prefix may be considered to be part of the same flow. The address of the next hop where
the packet should be forwarded to is the associated action. The length of the prefix defines
the priority of the rule.

2 Performance metrics for classification algorithms

• Search speed — Faster links require faster classification. For example, links run-
ning at 10Gbps can bring 31.25 million packets per second (assuming minimum
sized 40 byte TCP/IP packets).

• Low storage requirements — Small storage requirements enable the use of fast
memory technologies like SRAM (Static Random Access Memory). SRAM can
be used as an on-chip cache by a software algorithm and as on-chip SRAM for a
hardware algorithm.

• Ability to handle large real-life classifiers.

• Fast updates — As the classifier changes, the data structure needs to be updated.
We can categorize data structures into those which can add or delete entries incre-
mentally, and those which need to be reconstructed from scratch each time the
classifier changes. When the data structure is reconstructed from scratch, we call
it “pre-processing”. The update rate differs among different applications: a very
low update rate may be sufficient in firewalls where entries are added manually or
infrequently, whereas a router with per-flow queues may require very frequent
updates.

• Scalability in the number of header fields used for classification.

• Flexibility in specification — A classification algorithm should support general

rules, including prefixes, operators (range, less than, greater than, equal to, etc.)
and wildcards. In some applications, non-contiguous masks may be required.
 6

3 Classification algorithms

3.1 Background
For the next few sections, we will use the example classifier in Table 5 repeatedly. The
classifier has six rules in two fields labeled F1 and F2 ; each specification is a prefix of
maximum length 3 bits. We will refer to the classifier as C = { R j } and each rule R j as a
2-tuple: 〈 R j1, R j2〉 .
TABLE 5.

Rule F1 F2

R1 00* 00*

R2 0* 01*

R3 1* 0*

R4 00* 0*

R5 0* 1*

R6 * 1*

3.1.1 Bounds from Computational Geometry

There is a simple geometric interpretation of packet classification. While a prefix rep-

resents a contiguous interval on the number line, a two-dimensional rule represents a rect-
angle in two-dimensional euclidean space, and a rule in d dimensions represents a d -
dimensional hyper-rectangle. A classifier is therefore a collection of prioritized hyper-
rectangles, and a packet header represents a point in d dimensions. For example, Figure 3
shows the classifier in Table 5 geometrically in which high priority rules overlay lower
 7

111
P
110
R5 R6 101
100
011
R2
R3 010
R1 001
000
000 010 100 110
001 011 101 111
Figure 3 Geometric representation of the classifier in Table 5. A packet represents a point, for instance
P(011,110), in two-dimensional space. Note that R4 is hidden by R1 and R2.

priority rules. Classifying a packet is equivalent to finding the highest priority rectangle
that contains the point representing the packet. For example, point P(011,110) in Figure 3
would be classified by rule R 5 .

There are several standard geometry problems such as ray shooting, point location and
rectangle enclosure that resemble packet classification. Point location involves finding the
enclosing region of a point, given a set of non-overlapping regions. The best bounds for
point location in N rectangular regions and d > 3 dimensions are O(log N) time with
d d–1
O(N ) space;1 or O( ( log N ) ) time with O(N) space [7][8]. In packet classification,

hyper-rectangles can overlap making classification at least as hard as point location.

d
Hence, a solution is either impracticably large (with 100 rules and 4 fields, N space is
d–1
about 100MBytes) or too slow ( ( log N ) is about 350 memory accesses).

We can conclude that: (1) Multi-field classification is considerably more complex than
one-dimensional longest prefix matching, and (2) Complexity may require that practical
solutions use heuristics.

1. The time bound for d ≤ 3 is O(log log N) [7] but has large constant factors.
 8

3.1.2 Range Lookups

Packet classification is made yet more complex by the need to match on ranges as well
as prefixes. A range lookup for a dimension of width W bits can be defined as:

Definition 1: Given a set of N disjoint ranges G = { G i = [l i,u i] } that form a partition of

W
the number line [0,2 – 1] , i.e., l i and u i are such that
W
l 1 = 0, l i ≤ u i, l i + 1 = u i + 1, u N = 2 – 1 ; the range lookup problem is to find the
range G P (and any associated information) that contains an incoming point P .

To assess the increased complexity of ranges, we can convert each range to a set of
prefixes (a prefix of length s corresponds to a range [ l, u ] where the ( W – s ) least signif-
icant bits of l are all 0 and those of u are all 1) and use a longest prefix matching algo-
rithm [ref tutorial paper in same issue]. Table 6 shows some examples of range-to-prefix
conversions for W = 4 .
TABLE 6.

Range Constituent Prefixes

[4,7] 01**

[3,8] 0011, 01**, 1000

[1,14] 0001, 001*, 01**, 10**, 110*, 1110

A W -bit range can be represented by at most 2W – 2 prefixes (see the last row of Table
6 as an example) which means a prefix matching algorithm can find ranges with 2W times
as much storage. Feldman and Muthukrishnan [3] show a reduction of ranges to prefix
lookup with a two-fold storage increase that can be used in some specific multi-dimen-
sional classification schemes.
 9

3.2 Taxonomy of classification algorithms

The classification algorithms we will describe here can be categorized into the four
classes shown in Table 7.
TABLE 7.

Category Algorithms
Basic data Linear search, caching, hierarchical tries, set-pruning
structures tries

Geometry- Grid-of-tries, AQT, FIS

based

Heuristic RFC, hierarchical cuttings, tuple-space search

Hardware only Ternary CAM, bitmap-intersection

We now proceed to describe representative algorithms from each class.

3.3 Basic data structures

3.3.1 Linear search

The simplest data structure is a linked-list of rules stored in order of decreasing prior-
ity. A packet is compared with each rule sequentially until a rule is found that matches all
relevant fields. While simple and storage-efficient, this algorithm clearly has poor scaling
properties; the time to classify a packet grows linearly with the number of rules.

3.3.2 Hierarchical tries

A d -dimensional hierarchical radix trie is a simple extension of the one dimensional

radix trie data structure, and is constructed recursively as follows. If d is greater than 1,
we first construct a 1-dimensional trie, called the F1 -trie, on the set of prefixes { R j1 } ,
belonging to dimension F1 of all rules in the classifier, C = { R j } . For each prefix, p , in
the F1 -trie, we recursively construct a ( d – 1 ) -dimensional hierarchical trie, T p , on those
rules which specify exactly p in dimension F1 , i.e., on the set of rules { R j :R j1 = p } . Pre-
fix p is linked to the trie T p using a next-trie pointer. The storage complexity of the data
 10

search path
1
0
F1-trie
1

0 0 1 1 0
R3
0
R4 1
R5 R6 F2-tries
R1
R2
Figure 4 A hierarchical trie data structure. The gray pointers are the “next-trie” pointers. The path
traversed by the query algorithm on an incoming packet (000, 010) is shown.

structure for an N -rule classifier is O(NdW) . The data structure for the classifier in Table 5
is shown in Figure 4. Hierarchical tries are sometimes called “multi-level tries”, “back-
tracking-search tries”, or “trie-of-tries”.

Classification of an incoming packet ( v 1, v 2, …, v d ) proceeds as follows. The query

algorithm first traverses the F1 -trie based on the bits in v 1 . At each F1 -trie node encoun-
tered, the algorithm follows the next-trie pointer (if present) and traverses the ( d – 1 ) -
d
dimensional trie. The query time complexity for d -dimensions is therefore O(W ) . Incre-
2
mental updates can be carried out similarly in O(d W) time since each component of the
updated rule is stored in exactly one location at maximum depth O(dW) .

3.3.3 Set-pruning tries

A set-pruning trie data structure [12] is similar, but with reduced query time obtained
by replicating rules to eliminate recursive traversals. The data structure for the classifier in
Table 5 is shown in Figure 5. The query algorithm for an incoming packet ( v 1, v 2, …, v d )
need only traverse the F1 -trie to find the longest matching prefix of v1 , follow its next-
trie pointer (if present), traverse the F2 -trie to find the longest matching prefix of v1 , and
 11

search path F1-trie

1
0

1 0 1 F2-tries
0 R3
1 0 1
0 R4 R5 x R6 R6
1 R5 1 R5
R1 R2 R2
Figure 5 A set-pruning trie data structure. The gray pointers are the “next-trie” pointers. The path
traversed by the query algorithm on an incoming packet (000, 010) is shown.

so on for all dimensions. The rules are replicated to ensure that every matching rule will
be encountered in the path. The query time is reduced to O(dW) at the expense of
d d
increased storage of O(N dW) since a rule may need to be replicated O(N ) times. Update
d
complexity is O(N ) , and hence, this data structure works only for relatively static classifi-
ers.

3.4 Geometric algorithms

3.4.1 Grid-of-tries

The grid-of-tries data structure, proposed by Srinivasan et al [10] for 2-dimensional

classification, reduces storages space by allocating a rule to only one trie node as in a hier-
archical trie, and yet achieves O(W) query time by pre-computing and storing a switch
pointer in some trie nodes. A switch pointer is labeled with ‘0’ or ‘1’ and guides the
search process. The conditions which must be satisfied for a switch pointer labeled b ( b =
’0’ or ‘1’) to exist from a node w in the trie T w to a node x of another trie T x are (see Fig-
ure 6):
 12

s F1-trie
r
T
F2-tries
y
w x
Tw Tx
Figure 6 The conditions under which a switch pointer exists from node w to x.

1. T x and T w are distinct tries built on the prefix components of dimension F2 . T x

and T w are pointed to by two distinct nodes, say r and s respectively of the same
trie, T , built on prefix components of dimension F1 .
2. The bit-string that denotes the path from the root node to node w in trie T w con-
catenated with the bit b is identical to the bit-string that denotes the path from the
root node to node x in the trie T x .

3. Node w does not have a child pointer labeled b , and

4. Node s in trie T is the closest ancestor of node r that satisfies the above condi-
tions.

If the query algorithm traverses paths U1(s, root ( T x ) , y, x) and U2(r, root ( T w ) , w) in
a hierarchical trie, it need only traverse the path ( s, r, root ( T w ) , w, x ) on a grid-of-tries.
This is because paths U1 and U2 are identical (by condition 2 above) till U1 terminates
at node w because it has no child branch (by condition 3). The switch pointer eliminates
the need for backtracking in a hierarchical trie without the storage of a set-pruning trie.
Each bit of the packet header is examined at most once, so the time complexity reduces to
O(W) , while storage complexity O(NW) is the same as a 2-dimensional hierarchical trie.

However, switch pointers makes incremental updates difficult, so the authors [10] recom-
mend rebuilding the data structure (in time O(NW) ) for each update. An example of the
grid-of-tries data structure is shown in Figure 7.
 13

search path
1 F1-trie
0
1
0 0
R3
0 1 1 1 1

0 0
R4 1 F2-tries
1 R5 R6
R1
R2
Figure 7 The grid-of-tries data structure. The switch pointers are shown dashed. The path traversed by the
query algorithm on an incoming packet (000, 010) is shown.

Reference [10] reports 2MBytes of storage for a 20,000 two-dimensional classifier

with destination and source IP prefixes. The stride of the destination (source) prefix trie
was 8 (5) bits respectively, leading to a maximum of 9 memory accesses.

Grid-of-tries works well for two dimensional classification, and can be used for the
last two dimensions of a multi-dimensional hierarchical trie, decreasing the classification
d–1
time complexity by a factor of W to O(NW ) . As with hierarchical and set-pruning

tries, grid-of-tries handles range specifications by splitting into prefixes.

3.4.2 Cross-producting

Cross-producting [10] is suitable for an arbitrary number of dimensions. Packets are

classified by composing the results of separate 1-dimensional range lookups for each
dimension as explained below.

Constructing the data structure involves computing a set of ranges, G k , of size

j
s k = G k , projected by rule specifications in each dimension k, 1 ≤ k ≤ d . Let r k ,
 14

r11 r12 r13

Crossproduct Table
111 (r11,r21)
P R1
110 (r11,r22) R2
r23 R5 R6 101 (r11,r23) R5
100 (r12,r21) —
011 (r12,r22)
R2 R2
r22 R3 010
R5
r21 R1 001 (r13,r21) R3
000 (r13,r22) R3
000 010 100 110 R5
001 011 101 111 (r13,r23)

Figure 8 The table produced by the crossproducting algorithm and its geometric representation.

∏ sk
th
1 ≤ j ≤ s k , denote the j range in G k . A cross-product table C T of size is con-
k=1
 i1 i2 id 
structed, and the best matching rule for each entry  r 1 , r 2 , …, r d  , 1 ≤ i k ≤ s k, 1 ≤ k ≤ d is
pre-computed and stored. Classifying a packet ( v 1, v 2, …, v d ) involves a range lookup in
ik i1 i2 id
each dimension k to identify the range r k containing point v k . The tuple 〈 r 1 , r 2 , …, r d 〉
is then found in the cross-product table C T which contains the pre-computed best match-
ing rule. Figure 8 shows an example.
d
Given that N prefixes leads to at most 2N – 2 ranges, s k ≤ 2N and C T is of size O(N ) .
The lookup time is O(dt RL) where t RL is the time complexity of finding a range in one
dimension. Because of its high worst case storage complexity, cross-producting is suitable
for very small classifiers. Reference [10] proposes using an on-demand cross-producting
scheme together with caching for classifiers bigger than 50 rules in five dimensions.
Updates require reconstruction of the cross-product table, and so cross-producting is suit-
able for relatively static classifiers.
 15

0 1
F1-trie
0

search path

000, 001, 011 010, 011, 100, 111 100, 111 000, 011
R1 R2 R5 R6 R3
R4
Figure 9 The data structure of Section 3.4.3 for the example classifier of Table 5. The search path for
example packet P(011, 110) resulting in R5 is also shown.

3.4.3 A 2-dimensional classification scheme [6]

Lakshman and Stiliadis [6] propose a 2-dimensional classification algorithm where

one dimension, say F1 , is restricted to have prefix specifications while the second dimen-
sion, F2 , is allowed to have arbitrary range specifications. The data structure first builds a
trie on the prefixes of dimension F1 , and then associates a set G w of non-overlapping
ranges to each trie node, w , that represents prefix p . These ranges are created by (possibly
overlapping) projections on dimension F2 of those rules, S w , that specify exactly p in
dimension F1 . A range lookup data structure (e.g., an array or a binary search tree) is then
constructed on G w and associated with trie node w . An example is shown in Figure 9.

Searching for point P (v 1,v 2) involves a range lookup in data structure G w for each trie
node, w , encountered. The search in G w returns the range containing v 2 , and hence the
best matching rule. The highest priority rule is selected from the rules { R w } for all trie
nodes encountered during the traversal.

The storage complexity is O(NW) because each rule is stored only once in the data
structure. Queries take O(W log N) time because an O(log N) range lookup is performed for
 16

NW (00) NE (10) 00 01 10 11

SW (01) SE (11)

Figure 10 A quadtree constructed by decomposition of two-dimensional space. Each decomposition

results in four quadrants.

every node encountered in the F1 -trie. This can be reduced to O(W + log N) using frac-
tional cascading [1], but that makes incremental updates impractical.

3.4.4 Area-based quadtree

The Area-based Quadtree (AQT) was proposed by Buddhikot et al [2] for two-dimen-
sional classification. AQT allows incremental updates whose complexity can be traded off
with query time by a tunable parameter. Each node of a quadtree [1] represents a two
dimensional space that is decomposed into four equal sized quadrants, each of which is
represented by a child node. The initial two dimensional space is recursively decomposed
into four equal-sized quadrants till each quadrant has at most one rule in it (Figure 10
shows an example of the decomposition). Rules are allocated to each node as follows. A
rule is said to cross a quadrant if it completely spans at least one dimension of the quad-
rant. For instance, rule R6 spans the quadrant represented by the root node in Figure 10,
while R5 does not. If we divide the 2-dimensional space into four quadrants, rule R5
crosses the north-west quadrant while rule R3 crosses the south-west quadrant. We call the
set of rules crossing the quadrant represented by a node in dimension k , the k -crossing fil-
ter set ( k -CFS) of that node.
 17

111
P
110
R5 R6 101 search path {R6}
100 00 10
01
011 {R2, R4} {R5}
R2 R3 {R3}
010 01
R1 001 {R1}
000
000 010 100 110
001 011 101 111
Figure 11 An AQT data structure. The path traversed by the query algorithm for an incoming packet
P(001, 010), yields R1 as the best matching rule.

Two instances of the same data structure are associated with each quadtree node —
each stores the rules in k -CFS ( k = 1, 2 ). Since rules in crossing filter sets span at least
one dimension, only the range specified in the other dimension need be stored. Queries
proceed two bits at a time by transposing one bit from each dimension, with two 1-dimen-
sional lookups being performed (one for each dimension on k -CFS) at each node. Figure
11 shows an example.

Reference [2] proposes an efficient update algorithm that, for N two-dimensional

rules, has O(NW) space complexity, O(αW) search time and O(α α N) update time, where
α is a tunable integer parameter.

3.4.5 Fat Inverted Segment tree (FIS-tree)

Feldman and Muthukrishnan [3] propose the Fat Inverted Segment tree (FIS-tree) for
two dimensional classification as a modification of a segment tree. A segment tree [1]
stores a set S of possibly overlapping line segments to answer queries such as finding the
highest priority line segment containing a given point. A segment tree is a balanced binary
search tree containing the end points of the line segments in S . Each node, w , represents a
range G w , leaves represent the original line segments in S , and parent nodes represent the
 18

R6
R5
R2
R4 R3
R1

000 001 011 111

[000,111]
{R6} {R1, R2, R4, R5} {R2, R5} {R3}
[000,001] [100,111]
{R2, R5} {R3} [010,011]
[000,011] [100,111]
{R6}
{R1, R4} [000,111]
[000,001] [010,011]

Segment Tree FIS-tree

Figure 12 The segment tree and the 2-level FIS-tree for the classifier of Table 5.

union of the ranges represented by their children. A line segment is allocated to a node w
if it contains G w but not G parent(w) . The highest priority line segment allocated to a node
is pre-computed and stored at the node. A query traverses the segment tree from the root,
calculating the highest priority of all the pre-computed segments encountered. Figure 12
shows an example segment tree.

An FIS-tree is a segment tree with two modifications: (1) The segment tree is com-
pressed (made “fat” by increasing the degree to more than two) in order to decrease its
depth and occupies a given number of levels l , and (2) Up-pointers from child to parent
nodes are used. The data structure for 2-dimensions consists of an FIS-tree on dimension
F1 and a range lookup data associated with each node. An instance of the range lookup

data structure associated with node w of the FIS-tree stores the ranges formed by the F2 -
projections of those classifier rules whose F1 -projections were allocated to w .
 19

A query for point P (v 1,v 2) first solves the range lookup problem on dimension F1 .
This returns a leaf node of the FIS-tree representing the range containing the point v 1 . The
query algorithm then follows the up-pointers from this leaf node towards the root node,
carrying out 1-dimensional range lookups at each node. The highest priority rule contain-
ing the given point is calculated at the end of the traversal.

Queries on an l -level FIS-tree have complexity O( ( l + 1 ) t RL) with storage complexity

1+1⁄l
O(ln ) , where t RL is the time for a 1-dimensional range lookup. Storage space can be

traded off with search time by varying l . Modifications to the FIS-tree are necessary to
support incremental updates — even then, it is easier to support inserts than deletes [3].
The static FIS-tree can be extended to multiple dimensions by building hierarchical FIS-
trees, but the bounds are similar to other methods studied earlier [3].

Measurements on real-life 2-dimensional classifiers are reported in [3] using the static
FIS-tree data structure. Queries took 15 or less memory operations with a two level tree,
4-60K rules and 5MBytes of storage. Large classifiers with one million 2-dimensional
rules required 3 levels, 18 memory accesses per query and 100MBytes of storage.

3.5 Heuristics
As we saw in Section 3.1.1, the packet classification problem is expensive to solve in
the worst-case — theoretical bounds state that solutions to multi-field classification either
require storage that is geometric, or a number of memory accesses that is polylogarithmic,
in the number of classification rules. We can expect that classifiers in real networks have
considerable structure and redundancy that might be exploited by a heuristic. That is the
motivation behind the algorithms described in this section.
 20

Simple One-step Classification

2S=2128 2T=212

Recursive Classification
2S=2128 264 224 2T=212

Phase 0 Phase 1 Phase 2 Phase 3

Figure 13 Showing the basic idea of Recursive Flow Classification. The reduction is carried out in
multiple phases, with a reduction in phase I being carried out recursively on the image of the phase I-1. The
example shows the mapping of 2 S bits to 2 T bits in 3 phases.

3.5.1 Recursive Flow Classification (RFC)

RFC [4] is a heuristic for packet classification on multiple fields. Classifying a packet
involves mapping S bits in the packet header to a T bit action identifier, where T = log N ,
S
T « S . A simple, but impractical method could pre-compute the action for each of the 2

different packet headers, yielding the action in one step. RFC attempts to perform the
same mapping over several phases, as shown in Figure 13; at each stage the algorithm
maps one set of values to a smaller set. In each phase a set of memories return a value
shorter (i.e., expressed in fewer bits) than the index of the memory access. The algorithm,
illustrated in Figure 14, operates as follows:

1. In the first phase, d fields of the packet header are split up into multiple chunks
that are used to index into multiple memories in parallel. The contents of each
memory are chosen so that the result of the lookup is narrower than the index.
2. In subsequent phases, memories are indexed using the results from earlier
phases.
3. In the final phase, the memory yields the action.

The algorithm requires construction of the contents of each memory, detailed in [4].
 21

Preprocessed Tables

indx

Packet

Phase 0 Phase 1 Phase 2 Phase 3

Figure 14 Packet flow in RFC.

Reference [4] reports that with real-life four-dimensional classifiers of up to 1700

rules, RFC appears practical for 10Gbps line rates in hardware and 2.5Gbps rates in soft-
ware. However, the storage space and pre-processing time grow rapidly for classifiers
larger than about 6000 rules. An optimization described in [4] reduces the storage require-
ment of a 15,000 four-field classifier to below 4MBytes.

3.5.2 Hierarchical Intelligent Cuttings (HiCuts)

HiCuts [5] partitions the multi-dimensional search space guided by heuristics that
exploit the structure of the classifier. Each query leads to a leaf node in the HiCuts tree,
which stores a small number of rules that can be searched sequentially to find the best
match. The characteristics of the decision tree (its depth, degree of each node, and the
 22

111 (8*8,X,4)
P Cut across X
110
R5 R6 101
100 (2*8,Y,2) R2 R3 R3
Cut across Y R5 R6 R6
011
R2
R3 010
R1 R5
R1 001
R2
000
000 010 100 110
001 011 101 111
Figure 15 A possible HiCuts tree for the example classifier in Table 5. Each ellipse in the tree denotes an
internal node v with a tuple (size of 2-dimensional space represented, dimension to cut across, number of
children). Each square is a leaf node which contains the actual classifier rules.

local search decision to be made at each node) are chosen while preprocessing the classi-
fier based on its characteristics (see [5] for the heuristics used).

Each node, v , of the tree represents a portion of the geometric search space. The root
node represents the complete d -dimensional space, which is partitioned into smaller geo-
metric sub-spaces, represented by its child nodes, by cutting across one of the d dimen-
sions. Each sub-space is recursively partitioned until no sub-space has more than B rules,
where B is a tunable parameter of the pre-processing algorithm. An example is shown in
Figure 15 for two dimensions with B = 2 .

Parameters of the HiCuts algorithm can be tuned to trade-off query time against stor-
age requirements. On 40 real-life four-dimensional classifiers with up to 1700 rules,
HiCuts requires less than 1 MByte of storage with a worst case query time of 20 memory

accesses, and supports fast updates.

 23

Rule Specification Tuple Tuple Hash Table Entries

R1 (00*,00*) (2,2) (0,1) {R6}
R2 (0**,01*) (1,2) (1,1) {R3,R5}
R3 (1**,0**) (1,1) (1,2) {R2}
R4 (00*,0**) (2,1) (2,1) {R4}
R5 (0**,1**) (1,1) (2,2) {R1}
R6 (***,1**) (0,1)
Figure 16 The tuples and associated hash tables in the tuple space search scheme for the example
classifier of Table 5.

3.5.3 Tuple Space Search

The basic tuple space search algorithm (Suri et al [11]) decomposes a classification
query into a number of exact match queries. The algorithm first maps each d -dimensional
th th
rule into a d -tuple whose i component stores the length of the prefix specified in the i
dimension of the rule (the scheme supports only prefix specifications). Hence, the set of
rules mapped to the same tuple are of a fixed and known length, and can be stored in a
hash table. Queries perform exact match operations on each of the hash tables correspond-
ing to all possible tuples in the classifier. An example is shown in Figure 16.

Query time is M hashed memory accesses, where M is the number of tuples in the
classifier. Storage complexity is O(N) since each rule is stored in exactly one hash table.
Incremental updates are supported and require just one hashed memory access to the
hashed table associated with the tuple of the modified rule. In summary, the tuple space
search algorithm performs well for multiple dimensions in the average case if the number
of tuples is small. However, the use of hashing makes the time complexity of searches and
d
updates non-deterministic. The number of tuples could be very large, up to O(W ) , in the
worst case. Furthermore, since the scheme supports only prefixes, the storage complexity
d
increases by a factor of O(W ) for generic rules as each range could be split into O(W)
prefixes in the manner explained in Section 3.1.2.
 24

Destination
Address

memory location: 1 2 N
Memory
Array TCAM
matched: 0 1 0 1

Priority
Encoder

memory
location
Action
RAM
Memory

Next-hop
Figure 17 The lookup operation using a ternary CAM.

3.6 Hardware-based algorithms

3.6.1 Ternary CAMs

A TCAM stores each W -bit field as a (val, mask) pair; where val and mask are each
W -bit numbers. For example, if W = 5 , a prefix 10* will be stored as the pair (10000,

11000). An element matches a given input key by checking if those bits of val for which
the mask bit is ‘1’, match those in the key.

A TCAM is used as shown in Figure 17. The TCAM memory array stores rules in
decreasing order of priorities, and compares an input key against every element in the
array in parallel. The N -bit bit-vector, matched, indicates which rules match and so the N -
bit priority encoder indicates the address of the highest priority match. The address is used
to index into a RAM to find the action associated with this prefix.
 25

TCAMs are being increasingly deployed because of their simplicity and speed (the
promise of single clock-cycle classification). Several companies produce 2Mb TCAMs
capable of single and multi-field classification in as little as 10ns. Both faster and denser
TCAMs can be expected in the near future. There are, however, some disadvantages to
TCAMs:

1. A TCAM is less dense than a RAM, storing fewer bits in the same chip area.
One bit in an SRAM typically requires 4-6 transistors, while one bit in a TCAM
requires 11-15 transistors [9]. A 2Mb TCAM running at 100 MHz costs about $70
today, while 8 Mb of SRAM running at 200 MHz costs about $30. Furthermore,
range specifications need to be split into multiple masks, reducing the number of
d
entries by up to ( 2W – 2 ) in the worst case. If only two 16-bit dimensions specify
ranges, this is a multiplicative factor of 900. Newer TCAMs, based on DRAM
technology, have been proposed and promise higher densities. One unresolved
issue with DRAM-based CAMs is the detection of soft errors caused by alpha par-
ticles.
2. TCAMs dissipate more power than RAM solutions because an address is com-
pared against every TCAM element in parallel. At the time of writing, a 2 Mb
TCAM chip running at 50 MHz dissipates about 7 watts [13][14]. In comparison,
an 8Mb SRAM running at 200 MHz dissipates approximately 2 watts [15].

TCAMs are appealing for relatively small classifiers, but will probably remain unsuit-
able in the near future for: (1) Large classifiers (256K-1M rules) used for microflow rec-
ognition at the edge of the network, (2) Large classifiers (128-256K rules) used at edge
routers that manage thousands of subscribers (with a few rules per subscriber), (3)
Extremely high speed (greater than 200Mpps) classification, and (4) Price-sensitive appli-
cations.

3.6.2 Bitmap-intersection

The bitmap-intersection classification scheme, proposed in [6], is based on the obser-

vation that the set of rules, S , that match a packet is the intersection of d sets, S i , where
th
S i is the set of rules that match the packet in the i dimension alone. While cross-pro-
 26

Dimension 1 Dimension 2
r11 {R1,R2,R4,R5,R6} 110111 r21 {R1,R3,R4} 101100
r12 {R2,R5,R6} 010011 r22 {R2,R3} 011000
r13 {R3,R6} 001001 r23 {R5,R6} 000111

Query on P(011,010): 010011 Dimension 1 bitmap

000111 Dimension 2 bitmap
000011
R5 Best matching rule
Figure 18 Bitmap tables used in the “bitmap-intersection” classification scheme. See Figure 8 for a
description of the ranges. Also shown is classification query on an example packet P(011, 110).

ducting pre-computes S and stores the best matching rule in S , this scheme computes S
and the best matching rule during each classification operation.

In order to compute intersection of sets in hardware, each set is encoded as an N -bit

bitmap with each bit corresponds to a rule. The set of matching rules is the set of rules
whose corresponding bits are ‘1’ in the bitmap. A query is similar to cross-producting:
First, a range lookup is performed in each of the d dimensions. Each lookup returns a bit-
map representing the matching rules (pre-computed for each range) in that dimension. The
d sets are intersected (a simple bit-wise AND operation) to give the set of matching rules,

from which the best matching rule is found. See Figure 18 for an example.

Since each bitmap is N bits wide, and there are O(N) ranges in each of d dimensions,
2
the storage space consumed is O(dN ) . Query time is O(dt RL + dN ⁄ w) where t RL is the
time to do one range lookup and w is the memory width. Time complexity can be reduced
by a factor of d by looking up each dimension independently in parallel. Incremental
updates are not supported.

Reference [6] reports that the scheme can support up to 512 rules with a 33 MHz field-
programmable gate array and five 1Mbit SRAMs, classifying 1Mpps. The scheme works
 27

well for a small number of rules in multiple dimensions, but suffers from a quadratic
increase in storage space and linear increase in classification time with the size of the clas-
sifier. A variation is described in [6] that decreases storage at the expense of increased
query time.

3.7 Summary of classification schemes

TABLE 8.

Worst-case time Worst-case storage

Algorithm
complexity complexity
Linear Search N N
Ternary CAM 1 N
Hierarchical Tries d
W NdW
Set-pruning Tries dW d
N
Grid-of-Tries d–1
W NdW
Cross-producting dW d
N
FIS-tree ( l + 1) W l×N
1+1⁄l

RFC d d
N
Bitmap-intersection dW + N ⁄ memwidth dN
2

HiCuts d d
N
Tuple Space Search N N

4 References
[1] M. de Berg, M. van Kreveld, and M. Overmars. Computational Geometry:
Algorithms and Applications, Springer-Verlag,2nd rev. ed. 2000.
 28

[2] M.M. Buddhikot, S. Suri, and M. Waldvogel. “Space decomposition tech-

niques for fast layer-4 switching,” Proceedings of Conference on Protocols
for High Speed Networks, pages 25-41, August 1999.

[3] A. Feldman and S. Muthukrishnan. “Tradeoffs for packet classification,”

Proceedings of Infocom, vol. 3, pages 1193-202, March 2000.

[4] Pankaj Gupta and Nick McKeown, Packet Classification on Multiple Fields,
Proc. Sigcomm, Computer Communication Review, vol. 29, no. 4, pp 147-
60, September 1999, Harvard University.

[5] Pankaj Gupta and Nick McKeown, Packet Classification using Hierarchical
Intelligent Cuttings , Proc. Hot Interconnects VII, August 99, Stanford.
This paper is also available in IEEE Micro, pp 34-41, vol. 20, no. 1, Janu-
ary/February 2000.

[6] T.V. Lakshman and D. Stiliadis. “High-Speed Policy-based Packet Forward-

ing Using Efficient Multi-dimensional Range Matching”, Proceedings of
ACM Sigcomm, pages 191-202, September 1998.

[7] M.H. Overmars and A.F. van der Stappen. “Range searching and point loca-
tion among fat objects,” Journal of Algorithms, vol. 21, no. 3, pages 629-
656, November 1996.

[8] F. Preparata and M. I. Shamos. Computational geometry : an introduction,

Springer-Verlag, 1985.

[9] F. Shafai, K.J. Schultz, G.F. R. Gibson, A.G. Bluschke and D.E. Somppi.
“Fully parallel 30-Mhz, 2.5 Mb CAM,” IEEE Journal of Solid-State Cir-
cuits, vol. 33, no. 11, November 1998.

[10] V. Srinivasan, S. Suri, G. Varghese, and M. Waldvogel. “Fast and Scalable

Layer four Switching,” Proceedings of ACM Sigcomm, pages 203-14, Sep-
tember 1998.

[11] V. Srinivasan, S. Suri, and G. Varghese. “Packet Classification using Tuple

Space Search”, Proceedings of ACM Sigcomm, pages 135-46, September
1999.
 29

[12] P. Tsuchiya. “A search algorithm for table entries with non-contiguous

wildcarding,” unpublished report, Bellcore.

[13] Netlogic microsystems, at http://www.netlogicmicro.com. CIDR products

at http://www.netlogicmicro.com/products/cidr/cidr.html

[14] Sibercore Technologies, at http://www.sibercore.com

[15] ZBT-Datasheet docs from IDT, at http://www.idt.com/products/pages/

ZBT_DS_p.html.