SPIRE2015-Improved Practical Compact Dynamic Tries

Improved Practical Compact Dynamic Tries
Andreas Poyias1
and Rajeev Raman1
University of Leicester, UK. Email: {ap480,r.raman}@le.ac.uk
Abstract. We consider the problem of implementing a dynamic trie
with an emphasis on good practical performance. For a trie with n nodes
with an alphabet of size σ, the information-theoretic lower bound is
n log σ + O(n) bits. The Bonsai data structure [1] supports trie opera-
tions in O(1) expected time (based on assumptions about the behaviour
of hash functions). While its practical speed performance is excellent, its
space usage of (1 + )n(log σ + O(log log n)) bits, where is any constant
> 0, is not asymptotically optimal. We propose an alternative, m-Bonsai,
that uses (1 + )n(log σ + O(1)) bits in expectation, and supports oper-
ations in O(1) expected time (again based on assumptions about the
behaviour of hash functions). We give a heuristic implementation of m-
Bonsai which uses considerably less memory and is slightly faster than
the original Bonsai.
1 Introduction
In this paper, we consider practical approaches to the problem of implementing
a dynamic trie in a highly space-eﬃcient manner. A dynamic trie (also known
as a dynamic cardinal tree [2]) is a rooted tree, where each child of a node is
labelled with a distinct symbol from an alphabet Σ = {0, . . . , σ−1}. We consider
dynamic tries that support the following operations:
create(): Create a new empty tree.
getRoot(): return the root of the current tree.
getChild(v, i): return child node of node v with symbol i, if any (and return
−1 if no such child exists).
addChild(v, i): add new child with symbol i and return the newly created node.
getParent(v): return the parent of node v.
We do not discuss deletions explicitly, but do indicate what is possible with re-
gards to deletions. A trie is a classic data structure (the name dates back to 1959)
and has numerous applications in string processing. A naive implementation of
tries uses pointers. Using this approach, each node in an n-node binary trie uses
3 pointers for the navigational operations. A popular alternative for larger al-
phabets is the ternary search tree (TST) [3], which uses 4 pointers (3 plus a
parent pointer), in addition to the space for a symbol. Other approaches include
the double-array trie (DAT), which uses a minimum of two integers per node,
each of magnitude O(n). Since a pointer must asymptotically use Ω(log n) bits

of memory, the asymptotic space bound of TST (or DAT) is O(n(log n + log σ))
bits. However, the information-theoretic space lower bound of n log σ+O(n) bits
(see e.g. [2]) corresponds to one symbol and O(1) bits per node. Clearly, if σ is
small, both TST and DAT are asymptotically non-optimal. In practice, log σ
is a few bits, or one or two bytes at most. An overhead of 4 pointers, or 32n
bytes on today’s machines, makes it impossible to hold tries with even moder-
ately many nodes in main memory. Although tries can be path-compressed by
deleting nodes with just one child and storing paths explicitly, this approach (or
even more elaborate ones like [4]) cannot guarantee a small space bound.
Motivated by this, a number of space-efficient solutions were proposed [5, 2,
6–8], which represent static tries in information-theoretically optimal space, and
support a wide range of operations. A number of asymptotic worst-case results
were given in [9–12]. As our focus is on practical performance, we do not discuss
all previous results in detail and refer the reader to e.g. [11] for a comparison. For
completeness, we give a summary of some the results of [11, 12]. The first uses
almost optimal 2n+n log σ+o(n log σ) bits, and supports trie operations in O(1)
time if σ = polylog(n) and in O(log σ/ log log σ) time otherwise. The second [12,
Theorem 2] uses O(n log σ) bits and supports individual dynamic trie operations
in O(log log n) amortized expected time, although finding the longest prefix of a
string in the trie can be done in O(log k/ logσ n+log log n) expected time. Neither
of these has been fully implemented, although a preliminary attempt (without
memory usage measurements) was presented in [13]. Finally, we mention the
wavelet trie [14] which is a data structure for a sequence of strings, and in
principle can replace tries in many applications. Although in theory it is dynamic,
we are not aware of any implementation of a dynamic wavelet trie.
Predating most of this work, Darragh et al. [1] proposed the Bonsai data
structure, which uses a different approach to support the above dynamic trie
operations in O(1) expected time (based on assumptions about the behaviour of
hash functions). While its practical speed performance is excellent, we note here
that the asymptotic space usage of the Bonsai data structure is (1 + )n(log σ +
O(log log n)) bits, where is any constant > 0, which is not asymptotically
optimal due to the addition of O(log log n) term. The additive O(n log log n)
bits term can be significant in many practical applications where the alphabet
size is relatively small, including one involving mining frequent patterns that we
are considering. The Bonsai data structure also has a certain chance of failure:
if it fails then the data structure may need to be rebuilt, and its not clear how
to do this without affecting the space and time complexities.
In this paper, we introduce m-Bonsai1
, a variant of Bonsai. Again, based upon
the same assumptions about the behaviour of [1], our variant uses (1+ )n(log σ+
O(1)) bits of memory in expectation, where is any constant > 0, which is
asymptotically optimal, and operations take O(1) expected time. We give two
practical variants of m-Bonsai: m-Bonsai (γ) and m-Bonsai (recursive). Our
implementations and experimental evaluations show that m-Bonsai (recursive)
is consistently a bit faster than the original Bonsai and significantly more space-
1
This could be read as mame-bonsai, a kind of small bonsai plant, or mini-bonsai.

efficient than the original, while m-Bonsai (γ) is even more space efficient but
rather slower. Of course, all Bonsai variants use at least 20 times less space
than TSTs for small alphabets and compare well in terms of speed with TSTs.
We also note that our experiments show that the hash functions used in Bonsai
appear to behave in line with the assumptions about their behaviour. Finally, for
both Bonsai and m-Bonsai, we believe it is relatively easy to remove the (1 + )
multiplicative factor from the n log σ term, but since this is not our primary
interest is robust practical performance, we have not pursued this avenue.
The rest of this paper is organized as follows. In Section 2, we talk about the
asymptotics of Bonsai [1] and give a practical analysis. Section 3 summarizes
m-Bonsai approach which is followed by Section 4 the experimental evaluation.
2 Preliminaries
Bit-vectors. Given a bit string x1, . . . , xn, we define the following operations:
select1(x, i): Given an index i, return the location of ith 1 in x.
rank1(x, i): Return the number of 1s upto and including location i in x.
Lemma 1 (Pˇatra¸scu [15]). A bit string can be represented in n+O(n/(log n)2
)
bits such that select1 and rank1 can be supported in O(1) time.
Asymptotics of Bonsai We now sketch the Bonsai data structure, focussing
on asymptotics. It uses an array Q of size M to store a tree with n = αM
nodes for some 0 < α < 1 (we assume that n and M are known at the start
of the algorithm). We refer to α as the load factor. The Bonsai data structure
refers to nodes via a unique node ID, which is a pair i, j where 0 ≤ i < M
and 0 ≤ j < λ, where λ is an integer parameter that we discuss in greater
detail below. If we wish to add a child w with symbol c ∈ Σ to a node v with
node ID i, j , then w’s node ID is obtained as follows: We create the key of w
using the node ID of v, which is a triple i, j, c . We evaluate a hash function
h : {0, . . . , M · λ · σ − 1} → {0, . . . , M − 1} on the key of w. If i = h( i, j, c ),
the node ID of w is i , j where j ≥ 0 is the lowest integer such that there is
no existing node with a node ID i , j ; i is called the initial address of w.
In order to check if a node has a child with symbol c, keys are stored in Q using
open addressing and linear probing2
. The space usage of Q is kept low by the use
of quotienting [16]. The hash function has the form h(x) = (ax mod p) mod M
for some prime p > M · λ · σ and multiplier a, 1 ≤ a ≤ p − 1. Q only contains
the quotient value q(x) = (ax mod p)/M corresponding to x. Given h(x) and
q(x), it is possible to reconstruct x to check for membership. While checking
for membership for x, one needs to know h(y) for all keys y encountered during
the search, which is not obvious since keys may not be stored at their initial
address due to collisions. The Bonsai approach is to keep all keys with the same
initial address in consecutive locations in Q (this means that keys may be moved
2
A variant, bidirectional probing, is used in [1], but we simplify this to linear probing

after they have been inserted) and to use two bit-vectors of size M bits to effect
the mapping from a node’s initial address to the position in Q containing its
quotient, for details see [1]. Clearly, being able to search for, and insert keys
allows us to support getChild and addChild; for getParent(v) note that the key
of v encodes the node ID of its parent.
Asymptotic space usage. In addition to the two bit-vectors of M bits each, the
main space usage of the Bonsai structure is Q. Since a prime p can be found
that is < 2 · M · λ · σ, it follows that the values in Q are at most log2(2σλ + 1)
bits. The space usage of Bonsai is therefore M(log σ + log λ + O(1)) bits.
Since the choice of the prime p depends on λ, λ must be fixed in advance.
However, if more than λ keys are hashed to any value in {0, . . . , M − 1}, the
algorithm is unable to continue3
. Thus, λ should be chosen large enough to re-
duce the probability of more than λ keys hashing to the same initial address to
acceptable levels. In [1] the authors, assuming the hash function has full indepen-
dence and is uniformly random, argue that choosing λ = O(log M/ log log M)
reduces the probability of error to at most M−c
for any constant c (choosing
asymptotically smaller λ causes the algorithm almost certainly to fail). As the
optimal space usage for an n-node trie on an alphabet of size σ is O(n log σ)
bits, the additive term of O(M log λ) = O(n log log n) makes the space usage of
Bonsai non-optimal for small alphabets.
However, even this choice of λ is not well-justified from a formal perspective,
since the hash function used is quite weak—it is only 2-universal [17]. For 2-
universal hash functions, the maximum number of collisions can only be bounded
to O(
√
n) [18] (note that it is not obvious how to use more robust hash functions,
since quotienting may not be possible). Choosing λ to be this large would make
the space usage of the Bonsai structure asymptotically uninteresting.
Practical Analysis. In practice, we note that choosing λ = 32, and assuming
complete independence in the hash function, the error probability for M up to
264
is about 10−19
for α = 0.8, using the formula in [1]. Choosing λ = 16 as
suggested in [1] suggests a high failure probability for M = 256
and α = 0.8. Also,
in practice, the prime p is not significantly larger than Mλσ [19, Lemma 5.1].
The space usage of the Bonsai structure therefore is taken to be ( log σ + 7)M
bits for the tree sizes under consideration in this paper.
3 m-Bonsai
3.1 Overview
In our approach, each node again has an associated key that needs to be searched
for in a hash table, again implemented using open addressing with linear prob-
ing and quotienting. However, the ID of a node x in our case is a number from
3
Particularly for non-constant alphabets, it is not clear how to rebuild the data struc-
ture without an asymptotic penalty.

{0, . . . , M − 1} that refers to the index in Q that contains the quotient cor-
responding to x. If a node with ID i has a child with symbol c ∈ Σ, the
child’s key, which is i, c , is hashed using a multiplicative hash function h :
{0, . . . , M · σ − 1} → {0, . . . , M − 1}, and an initial address i is computed. If i
is the smallest index ≥ i such that Q[i ] is vacant, then we store q(x) in Q[i ].
Observe that q(x) ≤ 2σ , so Q takes M log σ +O(M) bits. In addition, we have
a displacement array D, and set D[i ] = i − i . From the pair Q[l] and D[l], we
can obtain both the initial hash address of the key stored there and its quotient,
and thus reconstruct the key. The key idea is that in expectation, the average
value in D is small:
Proposition 1. Assuming h is fully independent and uniformly random, the
expected value of
M−1
i=0 D[i] after all n = αM nodes have been inserted is ≈
M · α2
2(1−α) .
Proof. The average number of probes, over all keys in the table, made in a
successful search is ≈ 1
2 (1 + 1
1−α ) [16]. Multiplying this by n = αM gives the
total average number of probes. However, the number of probes for a key is
one more than its displacement value. Subtracting αM from the above and
simplifying gives the result.
Thus, encoding D using variable-length encoding could be very beneﬁcial. For
example, coding D in unary would take M +
M
i=1 D[i] bits; by Proposition 1,
and plugging in α = 0.8, the expected space usage of D, encoded in unary,
should be about 2.6M bits, which is smaller than the overhead of 7M bits of
the original Bonsai. As shown in Table 1, predictions made using Proposition 1
are generally quite accurate. Table 1 also suggests that encoding each D[i] using
the γ-code, we would come down to about 2.1M bits for the D, for α = 0.8.
unary γ Golomb
Load Factor 0.7 0.8 0.9 0.7 0.8 0.9 0.7 0.8 0.9
Pumsb 1.81 2.58 5.05 1.74 2.11 2.65 2.32 2.69 3.64
Accidents 1.81 2.58 5.06 1.74 2.11 2.69 2.33 2.69 3.91
Webdocs 1.82 2.61 5.05 1.75 2.11 2.70 2.33 2.70 3.92
Table 1. Average number of bits per entry needed to encode the displacement array
using the unary, γ and Golomb encodings. For the unary encoding, Proposition 1
predicts 1.81˙6, 2.6 and 5.05 bits per value. For ﬁle details see Table 2.
3.2 Representing the Displacement Array
We now describe how to represent the displacement array. A write-once dynamic
array is a data structure for a sequence of supports the following operations:
create(n): Create an array A of size n with all entries initialized to zero.
set(A, i, v): If A[i] = 0, set A[i] to v (assume 0 < v ≤ n). If A[i] = 0 then A[i]
is unchanged.
get(A, i): Return A[i].

The following lemma shows how to implement such a data structure. Note that
the apparently slow running time of set is enough to represent the displacement
array without asymptotic slowdown: setting D[i] = v means that O(v) time has
already been spent in the hash table finding an empty slot for the key.
Lemma 2. A write-once dynamic array A of size n containing non-negative
integers can be represented in space
n
i=1 |γ(A[i]+1)|+o(n) bits, supporting get
in O(1) time and set(A, i, v) in O(v) amortized time.
Proof. We divide A into contiguous blocks of size b = (log n)3/2
. The i-th block
Bi = A[bi..bi+b−1] will be stored in a contiguous sequence of memory locations.
There will be a pointer pointing to the start of Bi. Let Gi =
bi+b−1
j=bi |γ(A[j]+1)|.
We first give a naive representation of a block. All values in a block are
encoded using γ-codes and concatenated into a single bit-string (at least in
essence, see discussion of the get operation below). A set operation is performed
by decoding all the γ-codes in the block, and re-encoding the new sequence
of γ-codes. Since each γ-code is O(log n) bits, or O(1) words, long, it can be
decoded in O(1) time. Decoding and re-encoding an entire block therefore takes
O(b) time, which is also the time for the set operation. A get operation can
be realized in O(1) time using the standard idea of to concatenating the unary
and binary portions of the γ-codes separately into two bit-strings, and to use
select1 operations on the unary bit-string to obtain, in O(1) time, the binary
portion of the i-th γ-code. The space usage of the naive representation is i Gi+
O(( i Gi)/(log n)2
) + (n log n)/b) bits: the second term comes from Lemma 1
and the third accounts for the pointers and any unused space in the “last” word
of a block representation. This adds up to i Gi + o(n) bits, as required.
Since at most b set operations can be performed on a block, if any value
in a block is set to a value ≥ b2
, we can use the Ω(b2
) time allowed for this
operation to re-create the block in the naive representation, and also to amortize
the costs of all subsequent set operations on this block. Thus, we assume wlog
that all values in a block are < b2
, and hence, that γ-codes in a block are
O(log b) = O(log log n) bits long. We now explain how to deal with this case.
We divide each block into segments of = c log n/ log log n values for some
sufficiently small constant c > 0, which are followed by an overflow zone of at
most o =
√
log n log log n bits. Each segment is represented as a bit-string of
concatenated γ-codes. All segments, and their overflow zones, are concatenated
into a single bit-string. The bit-string of the i-th block, also denoted Bi, has
length at most Gi +(b/ )·o = Gi +O(log n(log log n)2
). As we can ensure that a
segment is of size at most (log n)/2 by choosing c small enough, we can decode an
individual γ-code in any segment in O(1) time using table lookup. We can also
support a set operation on a segment in O(1) time, by overwriting the sub-string
of Bi that represents this segment, provided the overflow zone is large enough
to accommodate the new segment.
If the overflow zone is exhausted, the time taken by the set operations that
have taken place in this segment alone is Ω(
√
log n log log n). Since the length
of Bi is at most O(
√
log n log log n) words, when any segment overflows, we

can simply copy Bi to a new sequence of memory words, and while copying,
use table lookup again to rewrite Bi, ensuring that each segment has an over-
flow zone of exactly o bits following it (note that as each segment is of length
Ω(log n/ log log n) bits and the overflow zones are much smaller, rewriting a col-
lection of segments that fit into O(log n) bits gives a new bit-string which is also
O(log n) bits).
One final component is that for each block, we need to be able to find the
start of individual segments. As the size of a segment and its overflow zone is
an integer of at most O(log log n) bits, and there are only O(
√
log n log log n)
segments in a block, we can store the sizes of the segments in a block in a single
word and perform the appropriate prefix sum operations in O(1) time using table
lookup, thereby also supporting get in O(1) time. This proves Lemma 2.
Theorem 1. For any given integers M and σ and constant 0 < α < 1, there is a
data structure that represents a trie on an alphabet of size σ with n nodes, where
n ≤ αM, using M log σ + O(M) bits of memory in expectation, and supporting
create() in O(M) time, getRoot and getParent in O(1) time, and addChild and
getChild in O(1) expected time. The expected time bounds are based upon the
assumption that the hash function has full randomness and independence.
Proof. Follows directly from Proposition 1 and Lemma 2, and from the obser-
vation that |γ(x + 1)| ≤ x + 2 for all x ≥ 0.
Remark 1. In the Bonsai (and m-Bonsai) approaches, deletion of an internal
node is in general not O(1) time, since the node IDs of all descendants of a node
are dependent on its own node ID. It is possible in m-Bonsai to delete a leaf,
taking care (as in standard linear probing) to indicate that a deleted location
in Q previously contained a value, and extending Lemma 2 to allow a reset(i)
operation, which changes A[i] from its previous value v to 0 in O(v) time.
3.3 Alternate Representation of the Displacement Array
The data structure of Lemma 2 appears to be too complex for implementation,
and a naive approach to representing the displacement array (as in Lemma 2)
may be slow. We therefore propose a practical alternative, which avoids any
explicit use of variable-length coding.
The displacement array is stored as an array D0 of fixed-length entries, with
each entry being ∆0 bits, for some integer parameter ∆0 ≥ 1. All displacement
values ≤ 2∆0
− 2 are stored as is in D0. If D[i] > 2∆0
− 2, then we set D0[i] =
2∆0
− 1, and store the value D [i] = D[i] − 2∆0
+ 1 as satellite data associated
with the key i in a second hash table.
This second hash table is represented using the original Bonsai representa-
tion, using a value M ∼ α n , where n is the number of keys stored in the
second hash table, and α is the load factor of this secondary hash table. The
satellite data for this second hash table are also stored in an array of size M
with fixed-length entries of size ∆1, where ∆1 is again an integer parameter. If

D [i] ≤ 2∆1
− 2, it is stored explicitly in the second-level hash table. Yet larger
values of D are stored in a standard hash table. The values of α , ∆0 and ∆1
are currently chosen experimentally, as described in the next section.
In what follows, we refer to m-Bonsai with the displacement array represented
as γ-codes as m-Bonsai (γ) and the representation discussed here as m-Bonsai
(recursive), respectively.
4 Experimental Evaluation
4.1 Implementation
We implemented m-Bonsai (recursive), m-Bonsai (γ) and Bonsai in C++, and
compared these with Bentley’s C++ TST implementation [3]. The DAT imple-
mentation of [20] was not tested since it apparently uses 32-bit integers, limiting
the maximum trie size to 232
nodes, which is not a limitation for the Bonsai or
TST approaches. The tests of [20] suggest that even with this “shortcut”, the
space usage is only a factor of 3 smaller than TST (albeit it is ∼ 2 times faster).
Both Bonsai implementations used the sdsl-lite library [21]. The origi-
nal Bonsai data structure mainly comprises three sdsl containers: firstly, the
int vector<>, which uses a fixed number of bits for each entry, is used for
the Q array (also in m-Bonsai). In addition, we use two bit vectors that to
distinguish nodes in collision groups as in [1]. In m-Bonsai (γ), D is split into
consecutive blocks of 256 values (initially all zero) each, which are stored as a
concatenation of their γ-codes. We used sdsl’s encode and decode functions to
encode and decode each block for the set and get operations.
The m-Bonsai (recursive) uses an alternative approach for the displacement
array. D0 has fixed length entries of ∆0-bits, thus int vector<> is the ideal
container. If a displacement value is larger than ∆0, we store it as a satellite data
in a Bonsai data structure. The satellite data is stored again in an int vector<>
of ∆1-bit entries. Finally, if the displacement value is even larger, then we use the
standard C++ std::map. In Figure 1, we show how we chose the parameters for
this approach. The three parameters α ,∆0 and ∆1 are selected given the trade-
off of runtime speed and memory usage. For this example we have α = 0.8.
Each line represents a different ∆0 value in bits. The y-axis shows the total bits
required per displacement value and the x-axis shows the choice of ∆1 sizes in
bits. As shown, there is a curve formed where its minimum point is when ∆1 = 7
for any ∆0 values. ∆0 = 3 is the parameter with the lower memory usage. ∆0 = 4
uses relatively more memory and even though ∆0 = 2 is closer to ∆0 = 3 in
terms of memory, it is slower in terms of runtime speed. This happens since less
values are accessed directly from D0 when ∆0 = 2, therefore we chose ∆0 = 3
and ∆1 = 7. Finally, we consider α = 0.8 as a good choice to have competitive
runtime speed and at the same time good memory usage.
4.2 Experimental Analysis
The machine used for the experimental analysis is an Intel Pentium 64-bit ma-
chine with 8GB of main memory and a G6950 CPU clocked at 2.80GHz with

Fig. 1. This graph is an example based on Webdocs8 used in m-Bonsai (recursive)
data structure with α = 0.8. The y-axis shows the bits per M in the displacement
array. The x-axis shows parameter ∆1 and each line is based on parameter ∆0.
3MB L2 cache, running Ubuntu 12.04.5 LTS Linux. All the code was compiled
using g++ 4.7.3 with optimization level 6. To measure the resident memory
(RES), /proc/self/stat was used. For the speed tests we measured wall clock
time using std::chrono::duration cast.
We use benchmark datasets arising arising in frequent pattern mining [22],
where each “string” is a subset of a large alphabet (up to tens of thousands). In
some frequent pattern mining algorithms such as [23], these strings need to be
traversed in sorted order, which takes a slow O(nσ) time in all Bonsai variants
because they do not support the next-sibling operation. To get around this, we
divide each symbol into 3-bit values, which optimizes the trade-off in runtime
speed and memory usage. Finally, we used sets of short read genome strings
given in the standard FASTQ format.
Memory Usage: For the memory usage experiments we set α = 0.8 for all Bonsai
data structures. Then, we insert all the strings of each dataset in the trees and
we measure resident memory. Table 2 is showing the average bits per n. It is
obvious that the Bonsai data structure is quite badly affected on datasets with
low σ. By converting the values of Table 2 in scale of bits per M ( as explained in
Section 3.1 n = αM), we prove the practical analysis of Section 2, showing that
Bonsai requires 10M-bits for the FASTQ sequences out of which 7M-bits are
used only to map the nodes in Q array. The m-Bonsai (γ) performance is very
good which needs more than 40% less memory than Bonsai on lower σ datasets.
The m-Bonsai (recursive) is also performing better than Bonsai and it is obvious
that as σ gets lower the recursive approach becomes more efficient by avoiding
the relatively big overhead of Bonsai.
Tree construction (Runtime speed): In Table 3 we show the wall clock time in
seconds for the construction of the Tree. The m-Bonsai (recursive) is proved
to be competitively fast and even faster than TST for some cases like Pumsb
and Accidents. This happens since m-Bonsai (recursive) is able to fit a big part

Datasets Node Number σ m-Bonsai (r) m-Bonsai(γ) Bonsai TST
Pumsb 1125375 7117 20.45 18.91 24 390.87
Accidents 4242318 442 15.65 14.12 19.2 388.26
Webdocs8 63985704 5577 20.45 18.91 24 386.79
Webdocs 231232676 5267657 27.04 30.91 36 385.1
splitPumsb 6702990 5 8.45 6.75 12 383.92
splitAccidents 17183926 5 8.45 6.75 12 387.07
splitWebdocs8 333444484 5 8.45 6.76 12
splitWebdocs 1448707216 5 8.45 6.78 12
SRR034939.fastq 3095560 5 8.45 6.73 12 385.88
SRR034944.fastq 21005059 5 8.45 6.76 12 385.76
SRR034940-1.fastq 1556235309 5 8.45 6.68 12
SRR034945-1.fastq 1728553810 5 8.45 6.68 12
Table 2. The average bits per node for datasets used for different purposes. In some
cases the TST processes were unable to finish execution due to large memory required.
of data structure in cache memory. However, when both data structures use
more heavily the main memory (Webdocs8), the pointer-based TST is 1.4 times
faster. The Bonsai implementation is consistently slower than TST and m-Bonsai
(recursive). Since the m-Bonsai (recursive) has a write once linear probing ap-
proach, when inserting a node in empty location Q[i], we know that D[i] is free
for insertions. Now, if D[i] is supposed to be zero then we don’t even need to
access D as it is already initialised to zeros4
. However, Bonsai always needs to
access at least one more bit-vector to reassure and mark the empty location.
Additionally, in case of collision Bonsai requires to swap elements in Q and one
of the bit-vectors, to make space for the new node at a matching location. Also,
if any satellite data(not included in this experiment) is required, it has to move
to match location as well thus potentially impacting the runtime performance.
Finally, the compact m-Bonsai (γ) is about ten times slower. This is due to the
O(b) time required to access each value as explained in Section 3.1 n = αM.
Datasets m-Bonsai (r) m-Bonsai(γ) Bonsai TST
Pumsb 0.55 5.97 0.86 0.64
Accidents 2.06 21.70 3.12 2.33
Webdocs8 27.03 255.25 35.13 19.38
Webdocs 110.35 886.17 125.06 608.91
splitPumsb 3.30 37.03 5.21 2.29
splitAccidents 7.72 82.95 10.92 5.69
splitWebdocs8 128.88 1287.49 173.25 1862.49
splitWebdocs 626.20 5439.71 832.8
SRR034939.fastq 0.561 9.82 0.74 0.61
SRR034944.fastq 6.041 72.38 6.84 4.39
SRR034940-1.fastq 746.005 5801.6 936.88
SRR034945-1.fastq 851.164 6456.18 1054.43
Table 3. The wall clock time in seconds for the construction of the Trie. Note that
the TST was affected by thrashing in Webdocs and splitWebdocs8.
4
Approximately 48% of the nodes have 0 displacement value at α = 0.8.

Successful search runtime speed: For this experiment we designed our own .search
datasets, where we randomly picked 10% of the strings from each dataset. As
shown in Table 4 we selected some datasets from our repository mainly due
to space limit. After the tree construction, we measured the time needed in
nanoseconds per successful search operation. It is obvious that TST is the fastest
approach. However, m-Bonsai (recursive) remains competitive with TST and
consistently faster than Bonsai by at least 1.5 times, whereas m-Bonsai (γ) in
the slowest. Note that there is an increase in runtime speed per search operation
for all Bonsai data structures as the datasets get bigger. However, we can’t prove
this for TST, since it is not able to process the larger datasets.
Datasets m-Bonsai (r) m-Bonsai(γ) Bonsai TST
Pumsb.search 237 1345 358 105
Webdocs8.search 332 1672 608 117
splitWebdocs.search 416 2037 657
SRR034940-1.search 403 1932 658
Table 4. The wall clock time in nanoseconds per successful search operations.
5 Conclusion
We have demonstrated a new variant of the Bonsai approach to store large tries in
a very space-efficient manner. Not only have we (re)-confirmed that the original
Bonsai approach is very fast and space-efficient on modern architectures, both
m-Bonsai variants we propose are significantly smaller (both asymptotically and
in practice) and and one of them is a bit faster than the original Bonsai. In the
near future we intend to investigate other variants, to give a less ad-hoc approach
to m-Bonsai (recursive), and to compare with other trie implementations.
Neither of our approaches is very close to the information-theoretic lower
bound of (σ log σ − (σ − 1) log(σ − 1))n − O(log(kn)) bits [2]. For example, for
σ = 5, the lower bound is 3.61n bits, while m-Bonsai (γ) takes ∼ 5.6M ∼ 7n bits.
Closing this gap would be an interesting future direction. Another interesting
open question is to obtain a practical compact dynamic trie that has a wider
range of operations, e.g. being able to navigate directly to the sibling of a node.
References
1. Darragh, J.J., Cleary, J.G., Witten, I.H.: Bonsai: a compact representation of trees.
Softw., Pract. Exper. 23(3) (1993) 277–291
2. Benoit, D., Demaine, E.D., Munro, J.I., Raman, R., Raman, V., Rao, S.S.: Rep-
resenting trees of higher degree. Algorithmica 43(4) (2005) 275–292
3. Bentley, J., Sedgewick, B.: Ternary search trees. http://www.drdobbs.com/
database/ternary-search-trees/184410528 (1998)
4. Nilsson, S., Tikkanen, M.: An experimental study of compression methods for
dynamic tries. Algorithmica 33(1) (2002) 19–33
5. Jacobson, G.: Space-efficient static trees and graphs. In: Proc. 30th Annual Sympo-
sium on Foundations of Computer Science, IEEE Computer Society (1989) 549–554

6. Raman, R., Raman, V., Satti, S.R.: Succinct indexable dictionaries with applica-
tions to encoding k-ary trees, prefix sums and multisets. ACM Transactions on
Algorithms 3(4) (2007)
7. Farzan, A., Munro, J.I.: A uniform paradigm to succinctly encode various families
of trees. Algorithmica 68(1) (2014) 16–40
8. Farzan, A., Raman, R., Rao, S.S.: Universal succinct representations of trees?
In Albers, S., Marchetti-Spaccamela, A., Matias, Y., Nikoletseas, S.E., Thomas,
W., eds.: Automata, Languages and Programming, 36th International Colloquium,
ICALP 2009, Rhodes, Greece, July 5-12, 2009, Proceedings, Part I. Volume 5555
of Lecture Notes in Computer Science., Springer (2009) 451–462
9. Munro, J.I., Raman, V., Storm, A.J.: Representing dynamic binary trees succinctly.
In Kosaraju, S.R., ed.: Proc. 12th Annual Symposium on Discrete Algorithms,
ACM/SIAM (2001) 529–536
10. Raman, R., Rao, S.S.: Succinct dynamic dictionaries and trees. In Baeten, J.C.M.,
Lenstra, J.K., Parrow, J., Woeginger, G.J., eds.: Automata, Languages and Pro-
gramming, 30th International Colloquium, ICALP 2003, Eindhoven, The Nether-
lands, June 30 - July 4, 2003. Proceedings. Volume 2719 of Lecture Notes in Com-
puter Science., Springer (2003) 357–368
11. Arroyuelo, D., Davoodi, P., Satti, S.: Succinct dynamic cardinal trees. Algorithmica
(2015) 1–36 Online first.
12. Jansson, J., Sadakane, K., Sung, W.: Linked dynamic tries with applications to
lz-compression in sublinear time and space. Algorithmica 71(4) (2015) 969–988
13. Takagi, T., Uemura, T., Inenaga, S., Sadakane, K., Arimura, H.: Applications of
succinct dynamic compact tries to some string problems. http://www-ikn.ist.
hokudai.ac.jp/~arim/papers/waac13takagi.pdf Presented at WAAC’13.
14. Grossi, R., Ottaviano, G.: The wavelet trie: maintaining an indexed sequence of
strings in compressed space. In: PODS. (2012) 203–214
15. Patrascu, M.: Succincter. In: 49th Annual IEEE Symp. Foundations of Computer
Science, IEEE Computer Society (2008) 305–313
16. Knuth, D.E.: The Art of Computer Programming, Volume 3: Sorting and Searching
(2nd ed.). Addison Wesley Longman (1998)
17. Carter, L., Wegman, M.N.: Universal classes of hash functions. J. Comput. Syst.
Sci. 18(2) (1979) 143–154
18. Fredman, M.L., Komlós, J., Szemerédi, E.: Storing a sparse table with 0(1) worst
case access time. J. ACM 31(3) (1984) 538–544
19. Pagh, R.: Low redundancy in static dictionaries with constant query time. SIAM
J. Comput. 31(2) (2001) 353–363
20. Yoshinaga, N., Kitsuregawa, M.: A self-adaptive classifier for efficient text-stream
processing. In: COLING 2014, 25th International Conference on Computational
Linguistics, Proceedings of the Conference: Technical Papers, August 23-29, 2014,
Dublin, Ireland. (2014) 1091–1102
21. Gog, S., Beller, T., Moffat, A., Petri, M.: From theory to practice: Plug and play
with succinct data structures. In: Experimental Algorithms - 13th International
Symposium, SEA 2014, Copenhagen, Denmark, June 29 - July 1, 2014. Proceed-
ings. (2014) 326–337
22. Goethals, B.: Frequent itemset mining implementations repository
http://fimi.ua.ac.be/.
23. Schlegel, B., Gemulla, R., Lehner, W.: Memory-efficient frequent-itemset mining.
In: EDBT 2011, 14th International Conference on Extending Database Technology,
Uppsala, Sweden, March 21-24, 2011, Proceedings. (2011) 461–472

SPIRE2015-Improved Practical Compact Dynamic Tries

More Related Content

What's hot (20)

Similar to SPIRE2015-Improved Practical Compact Dynamic Tries (20)

SPIRE2015-Improved Practical Compact Dynamic Tries