Engineering a Fast Online Persistent Suffix Tree Construction
Srikanta J. Bedathur
Database Systems Lab, SERC
Indian Institute of Science
Bangalore 560012, India
srikanta@dsl.serc.iisc.ernet.in
Jayant R. Haritsa
Database Systems Lab, SERC
Indian Institute of Science
Bangalore 560012, India
haritsa@dsl.serc.iisc.ernet.in
Abstract
Online persistent suffix tree construction has been considered
impractical due to its excessive I/O costs. However,
these prior studies have not taken into account the effects of
the buffer management policy and the internal node structure
of the suffix tree on I/O behavior of construction and
subsequent retrievals over the tree. In this paper, we study
these two issues in detail in the context of large genomic
DNA and Protein sequences. In particular, we make the following
contributions: (i) a novel, low-overhead buffering
policy called TOP-Q which improves the on-disk behavior
of suffix tree construction and subsequent retrievals, and (ii)
empirical evidence that the space efficient linked-list representation
of suffix tree nodes provides significantly inferior
performance when compared to the array representation.
These results demonstrate that a careful choice of
implementation strategies can make online persistent suffix
tree construction considerably more scalable – in terms
of length of sequences indexed with a fixed memory budget,
than currently perceived.
1. Introduction
The suffix tree is a versatile datastructure, that is used
in numerous bioinformatics applications (see [10, 11] for
a comprehensive list of these applications). For example,
they are extensively used in many tasks requiring similarity
searching over genetic sequences such as DNA and Proteins.
Although the utility of suffix trees is well known, their
usage is limited to small length datasets due to their space
requirements – the best in-memory implementations so far
take upto 12.5 times the database size [19]. This is in
marked contrast to traditional index structures (e.g., B+-
trees), where the size of the index is usually much smaller
than the indexed database.
Thus, it becomes untenable to consider a suffix tree residing
fully in memory, indexing an ever growing sequence
corpus such as the GenBank maintained by NCBI1. An obvious
solution to handle this space problem is to maintain
the suffix tree index on disk. Unfortunately, due to seemingly
random traversals induced by the linear-time construction
algorithms, resulting in unacceptably high I/O
costs, the folk wisdom is that disk based implementations
of suffix trees are unviable [8].
In order to overcome this infamous “memory bottleneck”
[13] of persistent suffix tree construction, there are
two possible approaches: (a) The suffix tree construction
algorithm and its structure could be modified to make it
more suitable for on-disk implementation, or (b) Tune the
parameters of the environment in which suffix tree is implemented,
without modifying either the structure or the construction
algorithm.
In a recent work, Hunt et al. [4], took the former approach
wherein they completely abandoned the use of suffix
links – additional edges over the internal nodes of the suffix
tree that are crucial in obtaining linear-time construction of
suffix tree. Using the resulting batch-wise construction with
quadratic worst case time complexity, they showed that suffix
trees can be built efficiently on-disk. However, due to
the resulting structure without suffix links, some of the fast
approximate string processing algorithms that make use of
suffix links, such as computing matching statistics [22], are
rendered unusable.
In this paper, we take the second approach, and identify
the parameters that affect online persistent suffix tree construction
and quantify their impact. Specifically, the contributions
of this paper are threefold:
1. We propose a novel paging strategy, TOP-Q, that takes
into account the probabilistic behavior of traversals
during suffix tree construction. This strategy uses only
the path length invariant (defined later) of suffix tree
1 As on January 2003, NCBI-GenBank houses about 28.5Gbp of DNA
sequence data [15]
ATTACT$
A
TTACT$
CT$
CT$
GTTAATTACT$
T
A TA
$
ATTACT$ CT$
ATTACT$
CT$
3
7
4
8
0
2 6
9
1 5
Figure 1. Suffix tree for the DNA fragment
“GTTAATTACT$”
nodes and results in a static policy, with extremely low
computational overhead.
2. We study a variety of paging policies in terms of their
utilization of buffer space during the persistent suffix
tree construction and compare their performance with
that of TOP-Q.
3. We show that the much preferred suffix tree implementation
using linked-list representation of nodes [19, 3],
is extremely expensive in terms of disk I/O in spite of
its space economy, in comparison to a simple and often
neglected array representation of edges.
In our experiments, we use the classical online suffix
tree construction algorithm of Ukkonen [7]. Our evaluation
testbed consists of a variety of real DNA sequences,
a synthetic symmetric Bernoulli sequence over a 4 character
alphabet, as well as a subset of the SWISS-PROT protein
database [20].
2. Persistent Suffix Tree Construction
In this section, we provide an overview of the online suffix
construction algorithm of Ukkonen, detailed in [10], and
consider the pattern of node accesses made during the algorithm’s
execution.
To aid the presentation, we first present some terminology
related to suffix trees. Let S[0::N] be the string indexed
by the tree TS. The leaf node corresponding to the i-th suffix,
S[i::N ], is represented as li. An internal node, v, has
an associated length L(v), which is the sum of edge lengths
on the path from root to v. We represent by (v), the string
at v to represent the substring S[i::i + L(v)] where li is
any leaf under v. A suffix link sl(v) = w exists for every
node v in the suffix tree such that if (v) = a, then
(w) = , where a is a single character of the alphabet
and  is a substring (possibly null) of the string. Note that
sl(v) is defined for every node in the suffix tree. And, more
importantly, sl(:) – the entire set of suffix links, forms a
tree rooted at the root of TS, with the depth of any node v
in this sl(:) tree being L(v). The suffix tree for an example
10-letter DNA subsequence, S = “GTTAATTACT$” is
shown in Figure 1. The numbers at the bottom of leaf nodes
represent the start of the suffix S[i::N ] that they represent.
The dashed edges between internal nodes represent the suffix
links.
The suffix tree is constructed incrementally by scanning
the string from left to right, one character at a time. A high
level description of the construction process is given in Algorithm
1. The algorithm can be viewed to consist of two
phases, Locate phase and Insert phase, for each character
in the sequence. If implemented naively, the locate-phase
would have quadratic time complexity, resulting in a overall
O(n3) time complexity. Therefore, the tree is augmented
with additional edges, called suffix links, that provide shortcuts
to move across the tree quickly. These suffix links play
a crucial role in reducing the running time of the algorithm
to linear in the length of the indexed string.
Algorithm 1 Online Algorithm of Ukkonen
1: procedure Ukkonen fOutputs an implicit suffix treeg
2: input S[0::m] : string to be indexed
3: I0 Implicit suffix tree for S[0 : : : 0]
4: for i = 0 to m do
5: for j = 0 to i + 1 do
6: fLOCATE PHASEg
7: Locate  = S[j : : : i] in Ii
8: fINSERT PHASEg
9: if  ends at a leaf then
10: Ii+1 add S[i + 1] to Ii
11: else f ends at an internal node, or the middle of
the edgeg
12: if from the end of  there is no path labeled
S[i + 1] then
13: Ii+1 split edge in Ii and add a new leaf
14: else
15: Ii+1 Ii f already exists in Iig
16: end if
17: end if
18: end for
19: end for
However, when we move the suffix tree construction
from memory to disk, these linear bounds no longer reflect
reality, since they were obtained with a RAM machine
model, where every memory access has the same cost, irrespective
of its address. On the other hand, access-cost in
secondary memory is dependent on the address to which
the previous access was made. For example, a long chain of
accesses to spatially contiguous addresses (block accesses)
could cost much less than fewer but random accesses.
Figure 2. Node Access Frequency
2.1. Node Access Patterns
During the construction of a suffix tree, accesses to nodes
are spatially non-contiguous. Specifically, in the locatephase,
already constructed parts of the tree are re-accessed
many times via suffix links. These traversals are not necessarily
spatially local, leading to seemingly random traversals
over the tree. The following words of Giegerich and
Kurtz [17], typifies the behavior of these algorithms: “The
active suffix creeps through the text like a caterpillar. At the
same time, the corresponding active node swings through
the tree like a butterfly”.
Thus, it is strongly believed that the accesses are random
in nature – with no obvious useful patterns discernible from
the access traces.
3. Locating Preferred Nodes
In this section, we closely analyse the traces of accesses
to nodes during online suffix tree construction, show that
some of the nodes are indeed accessed far more frequently
than others, and provide a simple observation that helps to
identify such nodes during construction of the tree.
Before we proceed to analyse the traces, we note that the
nature of accesses that are expected during the construction
of the suffix tree is intricately linked to the stochastic
properties of the specific sequence at hand. There have
been many efforts to classify the sequences based on their
stochastic properties [23]. One of the simplest sequence
models that is proposed as an approximation to genome sequences
is that of Bernoulli generators. In this model, symbols
of the alphabet are drawn independently of one another;
thus a string can be described as the outcome of a
sequence of Bernoulli trials. In addition, if all symbols are
drawn with equal probability, then the sequence is called
symmetric, otherwise, it is asymmetric.
Now, consider the internal node access statistics during
suffix tree construction for a symmetric Bernoulli sequence
(dataset S) derived from the access traces, shown in Figure
2. These provide the correlation between the average
number of accesses made to a node and the eventual depth
of the node in the tree, illustrating that, during the suffix tree
construction, nodes higher up in the tree are accessed more
number of times than nodes lower in the tree. This correlation
is also evident for suffix tree construction over real
chromosomal sequences (dataset C) as shown in Figure 2.
Thus, it seems reasonable to cache the nodes that end up
higher in the tree, in order to serve these accesses faster.
However, due to the nature of the edge splits during construction,
the depth of a node cannot be maintained without
propagating the update throughout the subtree under the
node.
3.1. Estimating the Depth of Internal Nodes
Although the depth of internal nodes cannot be maintained
accurately, a simple observation on the structure of
the resulting suffix tree provides us with a means to estimate
this value efficiently. Considering sequences drawn from a
symmetric Bernoulli stochastic models, it is straightforward
to see that:
 Substrings of equal length are equally likely.
 If s is a substring of the sequence S, with jsj = l,
then the number of substrings occurring elsewhere in
S which have a common prefix with s is directly proportional
to l.
Applying these to the behavior of the suffix tree during
its construction, we get:
Observation 1 The longer the edge in suffix tree of a symmetric
Bernoulli sequence, more the likelihood of its being
split in the limiting sense.
And obviously, no edge can be split once it has reached
the limiting minimum length of 1.
From the above observation, we can infer that a node,
in the limiting sense, can move further down in the suffix
tree until its incoming edge has only one symbol as its label.
Thus L(:) of any internal node forms an upper bound on
its eventual depth. In addition, this measure of path length
is an invariant for the node, which results in easy maintenance
of this information during the construction of the suffix
tree. Hence, for suffix tree nodes over large sequences,
we can approximate their eventual depth in the tree by their
path length.
3.2. Impact of Asymmetric Distribution
In deriving the above approximation to the eventual
depth of a node, we made the assumption that the sequence
is drawn from a symmetric Bernoulli model. However, it
is unlikely that any real-world DNA sequence would conform
to this restrictive model. The asymmetry of real-life
distribution results in substrings containing a larger proportion
of frequent symbols, having higher probability of occurrence
than other substrings of the same length. This implies
that if a node v has its label (v) containing smaller
number of frequent symbols, it has lower probability of being
split – resulting in a possible over-estimation of its eventual
depth by its path length L(v).
The graphs in Figure 3 show, for four DNA datasets used
in our experiments, the error in estimation of eventual depth
of a node vis-a-vis the actual depth of the node, as well as
the corresponding number of nodes at each depth of the tree.
These graphs were obtained after processing 5Mbp of each
of the datasets. The figure shows the statistics for nodes only
upto a depth of 20, since the number of the internal nodes
at depths greater than 20 is too small to make any impact
although the error in estimation of eventual depth is quite
large for such nodes.
The average error at each depth was computed as the
arithmetic mean of (L(:) �� Depth(:)) for all nodes at that
depth. Note that since L(:) forms an upperbound on the
depth of the node, this value is always positive. The following
points have to be noted with respect to these graphs:
 For all the datasets, path length corresponds exactly
to the depth of the node up to a certain value of actual
depth, which is dependent on the length of the sequence
processed. As shown in the graphs, this value
is 10, after processing 5Mbp of each of the datasets.
 The number of nodes peaks at a depth of 11, at which
point the error in the estimation of depth is small for
all the datasets.
 For the dataset S, estimates are accurate throughout
providing empirical evidence for the soundness of Observation
1.
 The worst estimation error are with dataset C, which
has a highly skewed distribution of basepairs. How-
0
20
40
60
80
100
0 5 10 15 20
Average Error in
Estimation
Actual Depth
(Number of Edges from the Root)
Error in Estimation of Depth
(Dataset Size = 5Mbp)
C
DHS
0
20
40
60
80
100
120
140
0 5 10 15 20
Number of Internal Nodes
(x 10,000)
Actual Depth
(Number of Edges from the Root)
C
DHS
Figure 3. Depth Estimation Error
ever, a majority of nodes in the tree occur within depth
14, where errors are not too large.
 The depth of a node approaches its path length with
the increase in length of the sequence indexed. Therefore,
the estimation errors continue to decrease as we
process sequences of greater length.
In summary, these graphs show that for even for datasets
that deviate from the symmetric Bernoulli model, the impact
of errors incurred by the proposed approximation to
eventual depth is not very significant, and the quality of approximation
improves with the increase in the length of the
sequence.
4. Design of TOP-Q
Before describing the TOP-Q buffering strategy, we
present the design of TOP, a simpler version of TOP-Q,
which exploits the observation that higher number of accesses
are to the nodes that are eventually higher in the
tree, as well as the approximation to the eventual depth presented
above.
We consider the situation where each disk-page contains
a collection of nodes of the suffix tree – either internal or
leaf, but not a mix of both. In order to minimize the storage
cost of maintaining the path length, each disk page contains
an associated path length, which is the average of the
path lengths of all nodes packed in it. Each disk-page is
completely packed with nodes as they created, and since
the path length for each node is an invariant, the path length
of the page can also be computed at the time it is committed
to the disk.
Using this depth estimation for each page, the ranking
policy employed for paging is to rank the pages with
smaller path lengths for retention in memory. We call this
buffering policy as TOP, to indicate that it tries to retain
those pages of the suffix tree that are estimated to be top
pages i.e., pages containing the top nodes of the tree.
4.1. Accomodating Correlated Accesses
Although the TOP buffering policy exploits the preferential
access to the nodes with lower path lengths, it ignores
the presence of correlated access patterns exhibited by the
suffix tree construction.We provide below an example situation
in the construction process, where this makes a impact
on the performance of TOP.
The construction proceeds by splitting an edge, introducing
a new branching node and a leaf node at that location,
and filling in suffix-link pointers if needed. Every node
stores the details of its incoming edge, i.e., (start; end) indexes
into the sequence and the length and label of the edge.
When an edge p(v) ! v, is about to be split, the following
actions are performed:
1. create a new branching node, v0, and a leaf node l,
2. set the incoming edge details for both v0 and l,
3. update the incoming edge details for v (the edge length
is shortened, and the start value and corresponding
edge label are changed), and finally,
4. v0 is set as the child of p(v) in place of v, and v is now
located under v0.
In addition, by the nature of the algorithm, only v has an active
reference to it, and p(v) is to be accessed through the
parent pointer available with v. Therefore, p(v) is not guaranteed
to be pinned in memory during this process.
As the construction progresses, the internal nodes have
ever-increasing path lengths associated with them. Therefore,
the TOP policy evicts the pages as soon as they are
filled and are not pinned through any active reference, since
the internal pages also have larger path length values as the
construction proceeds. Therefore, there is a possibility that
p(v), more precisely, the page containing p(v), is not retained
in the buffer for long.
We evaluated the impact of such correlated accesses to
nodes, by measuring the the number of characters processed
Figure 4. Correlated Accesses in TOP
by the algorithm between the instant a page is evicted and
the instant it is next requested, generating a fault on the
buffer pool. The initial interesting portion of the results is
shown in Figure 4, which plots the number of faulted pages
on a logscale and the number of characters processed since
they were last evicted from the buffer pool. As shown in
these graphs, the number of evictions that have pages with
immediate reference – i.e., before the complete addition of
the next character into the suffix tree – is orders of magnitude
larger than the evictions of pages that are accessed after
many more characters are added.
The TOP-Q strategy compensates for this unresponsiveness
of TOP to such accesses, by splitting
the buffer pool into a collection of pages maintained in the
order of their path lengths – implemented as a Heap structure,
and a short fixed-length queue of pages to hold the
pages evicted from the heap. The buffered pages in the
heap are chosen for eviction just like in the TOP policy.
However, unlike TOP, these pages are moved to the
short, fixed-length queue part of the buffer pool managed
in a FIFO fashion. The presence of the queue of pages
effectively introduces a delay in the eviction of pages, satisfying
almost all the immediate references to the page. We
have used a queue of 10 pages with buffer-pool sizes ranging
from thousands to hundreds of thousands of pages
in our experiments and found it to perform well in practice.
5. Suffix Tree Representation
We now turn our focus to the physical representation
of the nodes of the suffix tree. Much attention has been
paid to reducing the size of these nodes [19], the goal being
to maintain the tree entirely in memory, so that nonlocal
accesses over the tree induced by linear-time construction
algorithms are not affected by the virtual memory pagChar.
Label
Length of
Incoming Edge
Begin
Offset
Pointer to
Parent
(1 byte) (4 bytes) 4 bytes (4 bytes)
Suffix
Link
(4 bytes)
Child Pointers
(4 x 4 bytes)
Char.
Label
Length of
Incoming Edge
Begin
Offset
Pointer to
Parent
(1 byte) (4 bytes) 4 bytes (4 bytes)
Leaf Node Structure (=13 Bytes)
Internal Node Structure (=33 Bytes)
Char.
Label
Length of
Incoming Edge
Begin
Offset
Pointer to
Child
Pointer to
Next/Parent
Char.
Label
Length of
Incoming Edge
Begin
Offset
Pointer to
Next
Internal Node Structure (=21 Bytes)
(1 byte) (4 bytes) (4 bytes) (4 bytes) (4 bytes)
Suffix
Link
(4 bytes)
Leaf Node Structure (=13 Bytes)
(1 byte) (4 bytes) 4 bytes (4 bytes)
Array Based Representation Linked−list Representation
Figure 5. Structure of Suffix tree Nodes
ing [10]. However, it is not known which of these node representations
is more appropriate when the suffix tree has to
be constructed and maintained completely on disk.
The simplest but most space consuming strategy is to use
an array of size jj (where  is the size of the alphabet)
at each internal node of the tree. Each array entry corresponds
to an edge, whose edge-label begins with the character
associated with the array entry. These edges are implemented
as pointers to corresponding child nodes in the suffix
tree. We term this representation as array implementation
of suffix tree nodes. This representation, although simple
to program, is not preferred in most practical implementations
since this results in a lot of wasted space, with many
pointers containing null values. This overhead is especially
severe in nodes lower in the tree since the tree edges become
sparser at lower portions of the tree.
As suggested by McCreight [6], a space efficient alternative
to the array is to use a linked list of siblings and at every
internal node maintain a single pointer to the head of the
linked list containing its child nodes. Traversals from the internal
node are implemented by sequentially searching this
list for the appropriate child node. We refer to the resulting
representation as the linked-list representation of the suffix
tree node. This structure is a popular choice, due to its simplicity
in implementation as well as superior space economy.
The structure of internal nodes and leaf nodes, for the array
and linked-list implementation is given in Figure 5. It
is clear that the linked-list representation achieves its superior
space economy by significantly reducing the size of every
internal node – in our implementation, from 33 bytes
per internal node in array representation to 21 bytes per internal
node.
Although it is possible to maintain the linked-list of siblings
in a sorted fashion to reduce the searching time, this
provides no significant advantage with small alphabet size
sequences (such as DNA) and only complicates the implementation.
Moreover, maintaining the sorted order of the
list on disk involves update of many nodes for every insertion
leading to a large number of accesses. Hence, we do
Name Description
%-age Distribution
Length of nucleotides
(in Mbp) A T C G
S Sym. Bernoulli 25 25 25 25 25
D Drosoph. genome 25 29 29 21 21
H Human Chr. II 25 30 30 20 20
C C.elegans Chr. I 15 32 32 18 18
Table 1. Characteristics of the Datasets
not maintain the sibling linked-list in sorted fashion.
6. Evaluation Framework
In this section, we describe the framework used for evaluation
of various buffering strategies, and node implementation
choices, during the online construction of persistent
suffix trees.
In our evaluation, we use a total of four DNA datasets,
three of which are drawn from C.elegans Chromosome I
(dataset C), Human Chromosome II (dataset H) and complete
genome of Drosophila Melanogaster (dataset D). The
remaining dataset is a synthetic symmetric Bernoulli sequence
over DNA alphabet (dataset S). The details of these
datasets are summarized in Table 1. From these statistics we
see that these datasets comprise of sequences ranging from
symmetric distribution of alphabets (dataset S) to highly
skewed distribution (dataset C). Additionally, we also use
an amino-acid sequence, SPROT, derived from SWISSPROT
collection of proteins [20]. Since most of the individual
protein sequences are short, a 25Mb dataset was generated
by concatenating the sequences together.
6.1. Implementation Details
As already mentioned in Section 4, suffix tree nodes are
packed into fixed size pages before they are committed to
the disk. The pages on disk are either internal pages or leaf
pages, depending on whether they store internal nodes or
leaf nodes of the tree. The variation in the internal and leaf
node sizes leads to the packing density (i.e., the number of
nodes in a disk page) of leaf pages being greater than that
of internal pages.
The storage of both internal nodes and leaf nodes is in
their order of creation. Each page is committed immediately
to the disk, as soon as all the space in the page is utilized.
Note that, at the time of committing to the disk, all the entries
in nodes of the page may not be filled – some of them
may be defined and updated at later times in the construction
process. Also, a page is pinned in memory if there are
any active references pointing to nodes in the page.
The internal and leaf pages are distinguished also in
terms of their buffer pools. This is due to the distinctly different
access patterns made during the construction of the
tree. Internal nodes of the tree are used repeatedly (with or
without suffix links) for reaching the location of next suffix.
On the other hand, leaf nodes are re-visited only when the
suffix being located ends in a leaf node. In fact, our buffer
management system is designed such that separate policies
can be applied to internal and leaf page buffer pools.
6.2. Buffer Management Policies
The design of buffer management policies has been an
active area of research for many years, and a host of policies
that show improved hitrates over various database workloads
have been proposed [21, 2, 9, 5].
We compare the static policy of TOP-Q against the following
popular policies that are based on page access statistics,
commonly used in database management systems:
LRU (Least Recently Used) In case of a page fault, it replaces
the least recently used page from the buffer pool
to accommodate the new page. It incurs a constant time
computational overhead for every access, in order to
manipulate the list of pageframes maintained in the order
of recency of access.
2Q The 2Q algorithm [12] is a constant time overhead approximation
of LRU-2 [5] that is found to perform as
well as LRU-2 for a variety of reference patterns. The
2Q algorithm covers the most important drawback of
LRU-2, by reducing the computational overhead from
log(N) work for every access in LRU-2 to a constant
time overhead.
The metric of comparison between these popular buffering
policies and our TOP-Q strategy is the overall buffer
hit-rates observed during the construction of the suffix tree,
with increasing length of sequence indexed, and for a fixed
amount of buffer space.
6.3. Buffer Pool Allocation
In our evaluation of buffering policies, we maintain separate
buffer pools for leaf and internal pages, with the buffering
policy applied within each of the buffer pools. With a
fixed amount of memory space at our disposal as the buffer
space, it is interesting to see if there is an effective way to
partition this space between the two classes of pages of the
suffix tree.
The simplest partitioning is to distribute the available
buffer pages equally between both leaf and internal nodes.
It results in more number of leaf nodes being buffered than
the internal nodes due to the better packing density of the
leaf pages. Therefore, the effectiveness of buffer pool can
be improved by partitioning it to hold equal number of internal
nodes and leaf nodes. Although the number of internal
nodes is 0.6 - 0.8 times the number of leaf nodes
(for typical DNA sequences), the level of activity over internal
nodes, in terms of their accesses and updates, is much
higher than over leaf nodes. Hence, partitioning schemes
that are skewed to hold more number of internal pages than
leaf pages can be expected to perform better in practice.
Additionally, it should be noted that, as the construction
of the suffix tree progresses, the overall size of the tree increases,
leading to traversals over the tree covering a larger
number of pages. In fact, a point may arrive when the available
fixed size buffer may not be sufficient to efficiently
handle the requests over an extremely large suffix tree. In order
to compensate for this growing size of the data-structure
and provide a normalized performance measure for all the
policies, we consider the steady hitrates obtained, when a
fraction of the suffix tree size is provided for buffering. In
other words, as the suffix tree construction progresses, more
pages are introduced into the buffer pool such that the ratio
of buffer pool size to the total size of the suffix tree (measured
in number of pages) is held constant. The distribution
of steady hitrates obtained through these experiments
for various settings of fraction buffered, provide an upper
bound on the performance of each of the buffering policies,
independent of the size of the tree.
7. Experimental Results
In this section, we present the results of our empirical
evaluation of the buffering policies and the node implementation
choices outlined in previous sections. The parameters
applicable for all our experiments are summarized in
Table 2. Due to space limitations, we provide results for
only two DNA datasets, dataset S and dataset H, and for the
SPROT dataset. The results for other two DNA sequences,
dataset C and D, show behavior similar to that of dataset H
and S, respectively.
7.1. Construction with Fixed-size Buffer
The fixed buffer size experiments were conducted with
total memory space allocated for the buffer pool restricted
to just 32MB, a total of 8000 pages. This enabled us to work
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Equal Partition (4000/4000)
Dataset H
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Skewed Partition (7950/50)
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Equal Partition (4000/4000)
Dataset S
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Skewed Partition (7950/50)
TOP-Q
LRU
2Q
TOP
Figure 6. Combined Hit-rates for Construction with Array-based nodes
Parameter Value
Node size
Leaf 13 Bytes
Internal (Array) 33 Bytes
Internal (List) 21 Bytes
Page size 4096 Bytes
Table 2. Constant Experimental Parameters
with smaller length sequences, and perform experiments for
collecting buffer pool statistics using a simulated memory
hierarchy. However, in practice, much larger datasets will
be indexed and therefore proportionately larger buffer pool
sizes should be used.
As described earlier in Section 6.3, the available buffer
space can be partitioned between internal and leaf buffer
pools, ranging from equal partitioning to skewed partitioning
in favor of the internal buffer pool. We experimented
with many buffer partitioning schemes, and found
that the overall hitrate is dominated completely by the internal
node accesses alone. Thus, the performance improves
with increased skew in partitioning favoring the internal
pages. This observation holds for both the array as well as
linked-list representation of the suffix tree. In this paper, we
present results for equi-partitioning of the buffer pool with
4000 pages each for managing internal and leaf pages and
a highly skewed partitioning with 7950 pages to internal
buffer pool and remaining 50 pages as leaf buffers.
Array-based Suffix Tree Construction The buffer hitrate
obtained for page accesses, of both internal and leaf pages,
using array representation of nodes is shown in Figure 6.
The graphs also show the hitrate for simple TOP, in order to
provide a measure of gains obtained by the TOP-Q extension.
Figure 6 shows that TOP-Q provides consistently higher
hitrate than LRU and 2Q. In addition, the following observations
can be made about these results:
 The hitrates with skewed partitioning of buffer pool are
higher than with equi-partitioning. This clearly shows
that the overall performance is dominated by the effectiveness
of the buffering over internal pages.
 TOP-Q, as expected, performs better than the plain
TOP strategy and provides hitrates that degrade slower
with increasing suffix tree size, as compared to the
other policies.
 LRU exhibits the lowest hitrate, and upon further investigation
was found to have performance almost
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Equal Partition (4000/4000)
Dataset H
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Skewed Partition (7950/50)
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Equal Partition (4000/4000)
Dataset S
TOP-Q
LRU
2Q
TOP
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mbp)
Skewed Partition (7950/50)
TOP-Q
LRU
2Q
TOP
Figure 7. Combined Hit-rates for Construction with Linked-list Representation
equal to the policy of evicting a randomly selected
page (Random Replacement strategy).
 The performance of TOP is better than LRU in the initial
stages of the construction, and with increased skew
in the buffer allocation, TOP improves to match the
performance of 2Q.
 The ideal sequence for TOP is the dataset S, which
is a symmetric Bernoulli sequence. With this dataset,
it provides improved hitrates over even the highly sophisticated
2Q algorithm.
In summary, TOP-Q emerges as the best buffer management
policy with any buffer partitioning strategy over all
the datasets. Moreover, it should be noted that TOP-Q has
an extremely low computational overhead as compared to
LRU and 2Q, since its control data-structures are not updated
at every access to a page.
Linked-list based Suffix Tree Construction The behavior
of all the buffering policies for linked-list representation
of nodes is shown in Figure 7. First of all, it is worth noting
that all the algorithms provide better hitrates than in the
case of array representation, with the sole exception of TOP.
This is due to greater presence of correlated accesses, similar
to those discussed in Section 4.1, arising out of traversals
over the linked list of siblings. But the TOP-Q algorithm
still maintains an edge over the LRU and 2Q algorithms,
although the improvements are not as significant as
in the case of the array representation. Another interesting
point is that both LRU and 2Q show almost the same hitrates,
with both equal and skewed partitioning of the buffer
pool.
Choice of Implementation The comparison of graphs in
Figure 6 and Figure 7 indicates that the linked-list representation
provides better hitrates than array representation.
This is partly due to the fact that for the same amount of
buffer space at our disposal, the linked-list representation
buffers more number of internal nodes due to its improved
space economy. This seems to suggest that the linked-list
representation is better suited for persistent construction
with buffering.
However, this conclusion is misleading since the sequence
of page references in both cases are very different
and the hitrates are normalized within each reference sequence.
The absolute number of disk accesses made during
the construction provides a metric that is independent
of the reference sequence. Figure 8 shows the absolute number
of read and write disk accesses, using the TOP-Q buffering
policy, for the dataset H. As these numbers indicate, the
0
2e+07
4e+07
6e+07
8e+07
1e+08
1.2e+08
0 5 10 15 20 25 30
Number of Disk-accesses
Sequence Length (in Mbp)
Dataset H
Array
Linkedlist
Figure 8. Disk-accesses during construction
linked list representation has a significantly higher I/O overhead
than the array representation. This overhead is primarily
due to traversals over siblings in the linked-list to locate
the appropriate child to follow. Each of these siblings could
have been created at different points during the construction,
resulting in their non-contiguous storage on disk.
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
0 5 10 15 20 25 30
Hit Rate
Sequence Length (in Mb)
SPROT Dataset
TOP-Q, Skewed Partition
LRU, Skewed Partition
TOP-Q, Equal Partition
LRU, Equal Partition
Figure 9. Hitrates over Protein Data
Suffix Tree over Protein Data The results of our experiments
with the SPROT dataset are summarized in the Figure
9. The graph does not include hitrates for 2Q, since they
were very similar to LRU. As the numbers in this graph
demonstrate, TOP-Q performs better than LRU with increasing
skew in the buffer pool allocation. Further, with
increasing length of the indexed sequence, the hitrate of
TOP-Q degrades much slower than LRU, in both the partitioning
schemes. Thus, the benefits of TOP-Q are applicable
not only with small-alphabet sequences such as DNA,
but also with sequences with larger alphabet sizes.
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0 0.1 0.2 0.3 0.4 0.5 0.6
Steady Hitrate
Buffered Fraction
Dataset H
TOP-Q
LRU
2Q
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0 0.1 0.2 0.3 0.4 0.5 0.6
Steady Hitrate
Dataset S
TOP-Q
LRU
2Q
Figure 10. Behavior with Proportional Buffers
7.2. Construction with Proportional Buffering
We now move on to proportional buffering, discussed in
Section 6.3. The resulting steady hitrate values are plotted
in Figure 10.
As shown in these graphs, the performance of LRU improves
almost linearly with the buffered fraction of the
datastructure and 2Q is only marginally better than LRU.
On the otherhand, TOP-Q provides super-linear improvements
with diminishing returns, with increasing fraction of
buffering. The performance of TOP-Q peaks for about 25%
of the tree in the buffer providing close to 72.5% hitrate
for construction over real-life DNA sequences. When more
than 60% of the suffix tree is buffered, then all the three
buffering strategies perform equally well with upto 85%
steady hitrate.
7.3. On-disk Construction
In order to consider the practical impact of the improved
performance statistics for TOP-Q buffering and array representation
of nodes, we built persistent suffix trees for
large DNA sequences on two different classes of machines.
One was a PC class machine running Linux Redhat 8.0
Figure 11. Persistent Construction Times
and having an 18GB 10,000-RPM SCSI hard disk (IBM
DDYS-T18350M model). The other machine represents the
server class hardware used in current day large bioinformatics
projects – a HP-Compaq ES45 server running Tru64
Unix 5.1 and 432GB (6*72GB) storage with single channel
RAID controller at RAID-0 configuration. We refer to
these two platforms as PC and ES ,respectively, in the rest
of this discussion.
We compared the total execution time for constructing a
persistent suffix tree using 1GB of buffer space, split in a
moderately skewed fashion with 68% of pages to internal
buffer pool and the remaining 32% of pages as leaf buffer
pool. The strategies we used are: Linked-list representation
with LRU buffering, and array representation with TOP-Q
buffering scheme.
Figure 11 provides the performance of these two strategies
on both the platforms. These results show that TOP-Q
with array representation provides 50% to 65% improved
construction time over both platforms considered.
Impact of OS Buffering Due to the O SYNC option for
file operations used during the persistent suffix construction,
file writes are immediately synchronized with the disk
image. On the otherhand, the file reads might be served
from the I/O buffers maintained by the OS. Hence, the number
of disk page writes during construction makes a significant
impact on the overall construction time. Note that when
a page chosen as the victim by the buffering policy is dirty,
all the policies immediately flush it out to the disk. Upon
investigation we found that TOP-Q results in fewer diskwrites
than the other buffering strategies, which adds to the
reduction in seek-times achieved through improved hitrates,
translating to significant improvements in on-disk construction
times.
8. Related Work
Since the timeWeiner [16] introduced the suffix tree data
structure and a linear time algorithm for its construction,
there has been growing interest in more space- and timeefficient
algorithms for construction of suffix trees. A conceptually
different and space efficient algorithm to build
suffix trees in linear time have been given by McCreight [6],
and later, by Ukkonen [7]. Further, McCreight also suggested
the use of linked-list implementation of nodes for reducing
the space overhead of the suffix trees. All these algorithms
make use of suffix links to achieve linear-time construction
and are implemented for various constant-sized alphabet
datasets.
However, all of these algorithms show very bad performance
when applied to suffix tree construction in secondary
memory. The main bottleneck in the direct application
of these algorithms over suffix-trees on disk is considered
to be the random seeks induced during construction
[13]. It has also been noted in [4] that suffix links utilized
by all these algorithms traverse the suffix tree “horizontally”,
while edges span the tree “vertically”. Thus, atleast
one of them is expected to result in random access of
memory. However, this is true only if the tree is stored on
disk using depth-wise traversal of either edges of the tree or
links of the tree. But this storage pattern is not feasible during
the on-line construction of suffix tree on disk. Therefore,
in practice, both edges and suffix links show non-local
access patterns when the suffix tree is constructed on disk.
Thus, we need to carefully consider tuning of paging policies
to reduce the impact of such traversals.
Suffix trees provide an accurate indexing solution for
searching over large corpus of DNA or Protein sequences.
Initial use of suffix trees in genomic indexing was restricted
over small length DNA sequences [1], where suffix tree
could fit completely into main memory. However, until recently,
they were not considered for construction and maintenance
in secondary-memory. In fact, even in a recent
work [8], it was reported that whenever the dataset is large
enough suffix trees are not a viable option of indexing, since
the memory is too small to hold the index completely.
Initial theoretical breakthrough for suffix tree construction
on secondary memory was given by Farach et al. [13,
14]. They introduced a novel way to construct the suffix
tree over a large sequence by following a divide-conquer
approach (as opposed to the traditional suffix-at-a-time approach),
and used that to show that persistent suffix trees
could be built with the same I/O complexity as that of external
sorting. They also pointed out that “traditional” algorithms
(such as those of Weiner, Ukkonen, and McCreight
etc.), which follow incremental extension of suffix trees,
will be forced to make random I/Os resulting in bad ondisk
performance. Their observations on traditional algorithms
were made without considering the effects of paging/
caching policies that could be employed during the construction
process. In fact, they state at the end of their paper
[13], that it would be worthwhile considering the behavior
of the construction algorithms in presence of paging,
which is the topic we address in this paper.
Recently, Hunt et al. [4] proposed a phased construction
approach to building suffix trees, where they use an
asymptotically quadratic algorithm for construction of suffix
trees without suffix links. They report empirical evaluation
of both linear-time suffix link based algorithm and their
phased approach for persistent suffix trees to show the superiority
of their approach. Their results do not consider the
effects of paging policies and the storage management issues.
In fact, in [18], it has been reported that the bottleneck
is in the choice of checkpointing scheme of PJama – the underlying
persistent mechanism used in [4].
9. Conclusions
In this paper, we have evaluated the impact of buffering
and internal node implementation choices on the construction
of a suffix tree in secondary memory. We also
proposed a novel low overhead buffer management policy
called TOP-Q, which exploits the pattern of accesses over
the suffix tree during its construction. Through an extensive
empirical study involving both DNA and Protein sequences,
we showed that TOP-Q performs better than other popular
buffer algorithms such as LRU and LRU-2. The TOP-Q
algorithm saves more than 75% of disk I/O by buffering
merely 25% of the tree.
In addition, it was shown that that the commonly used,
space-economical linked-list representation of the suffix
tree is extremely expensive for construction on secondary
memory. Instead, a simple implementation using arrays at
each internal node is shown to be far better suited for persistent
suffix tree representation.
A performance evaluation of TOP-Q with array representation
of nodes against a popularly reported linked-list
representation with LRU buffering policy showed that significant
speedups to the tune of 50% to 70% were obtained,
over different platforms.
Acknowledgements This work was supported in part by
a Swarnajayanti Fellowship from the Dept. of Science &
Technology, Govt. of India, and a research grant from the
Dept. of Bio-technology, Govt. of India.
References
[1] P. Bieganski. Genetic sequence data retrieval and manipulation
based on generalized suffix trees. Ph. D. Thesis, University
of Minnesota, 1995.
[2] H. Chou and D. Dewitt. An evaluation of buffer management
strategies for relational database systems. In Proc. of
11th VLDB Conf., 1985.
[3] E. Hunt. PJama Stores and Suffix Tree Indexing for Bioinformatics
Applications. In Proc. of ECOOP, 2000.
[4] E. Hunt, M. P. Atkinson, and R.W. Irving. A Database Index
to Large Biological Sequences. In Proc. of the 27th VLDB
Conference, 2001.
[5] E. J. O’Neil, P. E. O’Neil, and G. Weikum. The lru-k page
replacement algorithm for database disk buffering. In Proc.
of the ACM SIGMOD Intl. Conf. on Mgmt. of Data, 1993.
[6] E. M. McCreight. A Space-Efficient Suffix Tree Construction
Algorithm. JACM, 23(2), 1976.
[7] E. Ukkonen. Online Construction of Suffix-trees. Algorithmica,
14, 1995.
[8] G. Navarro and R. Baeza-Yates. A Hybrid Indexing Method
for Approximate String Matching. Jl. of Discrete Algorithms,
2000.
[9] Giovanni Maria Sacco. Index Access with Finite Buffer. In
Proc. of 13th VLDB Conf., 1987.
[10] D. Gusfield. Algorithms on strings, trees and sequences:
computer science and computational biology. Cambridge
University Press, Cambridge, 1997.
[11] D. Gusfield. Suffix trees come of age in bioinformatics (invited
talk). In IEEE Bioinformatics Conference (CSB), 2002.
[12] T. Johnson and D. Shasha. 2Q : A Low Overhead High Performance
Buffer Management Replacement Algorithm. In
Proc. of the 20th VLDB Conf., 1994.
[13] M. Farach, P. Ferragina, and S. Muthukrishnan. Overcoming
the Memory Bottleneck in Suffix Tree Construction. Proc. of
IEEE Annual Symposium on Foundations of Computer Science,
1998.
[14] M. Farach-Colton. Optimal Suffix Tree Construction with
Large Alphabets. In Proc. of IEEE Annual Symposium on
Foundations of Computer Science, 1997.
[15] GenBank Overview. http://www.ncbi.nlm.nih.gov/Genbank/
GenbankOverview.html, 2002.
[16] P. Weiner. Linear Pattern Matching algorithms. In Proc.
of the 14th IEEE Symp. on Switching and Automata Theory,
1973.
[17] R. Giegerich and S. Kurtz. A Comparison of Imperative and
purely Functional Suffix tree constructions. Science of Programming,
1995.
[18] R. Japp. First Year Report. Master’s thesis, University of
Glasgow, July 2001.
[19] S. Kurtz. Reducing Space Requirement of Suffix Trees. Software
Practice and Experience, 29(13):1149–1171, 1999.
[20] Swiss-prot protein knowledgebase.
http://www.expasy.org/sprot/, 2002.
[21] W. Effelsberg and T. Haerder. Principles of Database Buffer
Management. ACM Transactions on Database Systems, 9(4),
1984.
[22] W. I. Chang and E. L. Lawler. Approximate string matching
in sublinear expected time. In Proc. of IEEE Annual Symposium
on Foundations of Computer Science, 1990.
[23] W. Szpankowski. Average Case Analysis of Algorithms on
Sequences. Wiley-Interscience, 2001.