BMZ.t2t

BMZ Algorithm


%!includeconf: CONFIG.t2t

----------------------------------------
==History==

At the end of 2003, professor [Nivio Ziviani http://www.dcc.ufmg.br/~nivio] was
finishing the second edition of his [book http://www.dcc.ufmg.br/algoritmos/].
During the [book http://www.dcc.ufmg.br/algoritmos/] writing, 
professor [Nivio Ziviani http://www.dcc.ufmg.br/~nivio] studied the problem of generating 
[minimal perfect hash functions concepts.html]
(if you are not familiarized with this problem, see [[1 #papers]][[2 #papers]]). 
Professor [Nivio Ziviani http://www.dcc.ufmg.br/~nivio] coded a modified version of 
the [CHM algorithm chm.html], which was proposed by
Czech, Havas and Majewski, and put it in his [book http://www.dcc.ufmg.br/algoritmos/].
The [CHM algorithm chm.html] is based on acyclic random graphs to generate 
[order preserving minimal perfect hash functions concepts.html] in linear time. 
Professor [Nivio Ziviani http://www.dcc.ufmg.br/~nivio] 
argued himself, why must the random graph 
be acyclic? In the modified version availalbe in his [book http://www.dcc.ufmg.br/algoritmos/] he got rid of this restriction.

The modification presented a problem, it was impossible to generate minimal perfect hash functions
for sets with more than 1000 keys.
At the same time, [Fabiano C. Botelho http://www.dcc.ufmg.br/~fbotelho],
a master degree student at [Departament of Computer Science http://www.dcc.ufmg.br] in 
[Federal University of Minas Gerais http://www.ufmg.br],
started to be advised by [Nivio Ziviani http://www.dcc.ufmg.br/~nivio] who presented the problem 
to [Fabiano http://www.dcc.ufmg.br/~fbotelho].

During the master, [Fabiano http://www.dcc.ufmg.br/~fbotelho] and 
[Nivio Ziviani http://www.dcc.ufmg.br/~nivio] faced lots of problems.
In april of 2004, [Fabiano http://www.dcc.ufmg.br/~fbotelho] was talking with a 
friend of him (David Menoti) about the problems
and many ideas appeared.
The ideas were implemented and a very fast algorithm to generate
minimal perfect hash functions had been designed.
We refer the algorithm to as **BMZ**, because it was conceived by Fabiano C. **B**otelho,
David **M**enoti and Nivio **Z**iviani. The algorithm is described in [[1 #papers]].
To analyse BMZ algorithm we needed some results from the random graph theory, so 
we invited professor [Yoshiharu Kohayakawa http://www.ime.usp.br/~yoshi] to help us.
The final description and analysis of BMZ algorithm is presented in [[2 #papers]].

----------------------------------------
 
==The Algorithm==

The BMZ algorithm shares several features with the [CHM algorithm chm.html].   
In particular, BMZ algorithm is also
based on the generation of random graphs [figs/img27.png], where [figs/img28.png] is in 
one-to-one correspondence with the key set [figs/img20.png] for which we wish to 
generate a [minimal perfect hash function concepts.html].
The two main differences between BMZ algorithm and CHM algorithm
are as follows: (//i//)  BMZ algorithm generates random 
graphs [figs/img27.png] with [figs/img29.png] and [figs/img30.png], where [figs/img31.png], 
and hence [figs/img32.png] necessarily contains cycles, 
while CHM algorithm generates //acyclic// random 
graphs [figs/img27.png] with [figs/img29.png] and [figs/img30.png],
with a greater number of vertices: [figs/img33.png];
(//ii//) CHM algorithm generates [order preserving minimal perfect hash functions concepts.html]
while BMZ algorithm does not preserve order.  Thus, BMZ algorithm improves
the space requirement at the expense of generating functions that are not
order preserving. 

Suppose [figs/img14.png] is a universe of //keys//.
Let [figs/img17.png] be a set of [figs/img8.png] keys from [figs/img14.png].
Let us show how the BMZ algorithm constructs a minimal perfect hash function [figs/img7.png].
We make use of two auxiliary random functions [figs/img41.png] and [figs/img55.png], 
where [figs/img56.png] for some suitably chosen integer [figs/img57.png], 
where [figs/img58.png].We build a random graph [figs/img59.png] on [figs/img60.png],
whose edge set is [figs/img61.png]. There is an edge in [figs/img32.png] for each 
key in the set of keys [figs/img20.png].

In what follows, we shall be interested in the //2-core// of
the random graph [figs/img32.png], that is, the maximal subgraph 
of [figs/img32.png] with minimal degree at 
least 2 (see [[2 #papers]] for details).
Because of its importance in our context, we call the 2-core the
//critical// subgraph of [figs/img32.png] and denote it by [figs/img63.png].
The vertices and edges in [figs/img63.png] are said to be //critical//.
We let [figs/img64.png] and [figs/img65.png].
Moreover, we let [figs/img66.png] be the set of //non-critical//
vertices in [figs/img32.png].
We also let [figs/img67.png] be the set of all critical
vertices that have at least one non-critical vertex as a neighbour.
Let [figs/img68.png] be the set of //non-critical// edges in [figs/img32.png].
Finally, we let [figs/img69.png] be the //non-critical// subgraph 
of [figs/img32.png].
The non-critical subgraph [figs/img70.png] corresponds to the //acyclic part//
of [figs/img32.png].
We have [figs/img71.png].

We then construct a suitable labelling [figs/img72.png] of the vertices
of [figs/img32.png]: we choose [figs/img73.png] for each [figs/img74.png] in such
a way that [figs/img75.png] ([figs/img18.png]) is a
minimal perfect hash function for [figs/img20.png].
This labelling [figs/img37.png] can be found in linear time
if the number of edges in [figs/img63.png] is at most [figs/img76.png] (see [[2 #papers]] 
for details).

Figure 1 presents a pseudo code for the BMZ algorithm.
The procedure BMZ ([figs/img20.png], [figs/img37.png]) receives as input the set of
keys [figs/img20.png] and produces the labelling [figs/img37.png].
The method uses a mapping, ordering and searching approach.
We now describe each step.
 | procedure BMZ ([figs/img20.png], [figs/img37.png])                              
 |     Mapping ([figs/img20.png], [figs/img32.png]);                
 |     Ordering ([figs/img32.png], [figs/img63.png], [figs/img70.png]); 
 |     Searching ([figs/img32.png], [figs/img63.png], [figs/img70.png], [figs/img37.png]);
 | **Figure 1**: Main steps of BMZ algorithm for constructing a minimal perfect hash function 

----------------------------------------

===Mapping Step===

The procedure Mapping ([figs/img20.png], [figs/img32.png]) receives as input the set 
of keys [figs/img20.png] and generates the random graph [figs/img59.png], by generating 
two auxiliary functions [figs/img41.png], [figs/img78.png].

The functions [figs/img41.png] and [figs/img42.png] are constructed as follows.
We impose some upper bound [figs/img79.png] on the lengths of the keys in [figs/img20.png].
To define [figs/img80.png] ([figs/img81.png], [figs/img62.png]), we generate 
an [figs/img82.png] table of random integers [figs/img83.png].
For a key [figs/img18.png] of length [figs/img84.png] and [figs/img85.png], we let 

 | [figs/img86.png]
 
The random graph [figs/img59.png] has vertex set [figs/img56.png] and 
edge set [figs/img61.png].  We need [figs/img32.png] to be 
simple, i.e., [figs/img32.png] should have neither loops nor multiple edges.
A loop occurs when [figs/img87.png] for some [figs/img18.png].
We solve this in an ad hoc manner: we simply let [figs/img88.png] in this case. 
If we still find a loop after this, we generate another pair [figs/img89.png].
When a multiple edge occurs we abort and generate a new pair [figs/img89.png]. 
Although the function above causes [collisions concepts.html] with probability //1/t//,
in [cmph library index.html] we use faster hash 
functions ([DJB2 hash http://www.cs.yorku.ca/~oz/hash.html], [FNV hash http://www.isthe.com/chongo/tech/comp/fnv/],
 [Jenkins hash http://burtleburtle.net/bob/hash/doobs.html] and [SDBM hash http://www.cs.yorku.ca/~oz/hash.html]) 
 in which we do not need to impose any upper bound [figs/img79.png] on the lengths of the keys in [figs/img20.png].

As mentioned before, for us to find  the labelling [figs/img72.png] of the 
vertices of [figs/img59.png] in linear time,
we require that [figs/img108.png]. 
The crucial step now is to determine the value 
of [figs/img1.png] (in [figs/img57.png]) to obtain a random 
graph [figs/img71.png] with [figs/img109.png].
Botelho, Menoti an Ziviani determinded emprically in [[1 #papers]] that 
the value of [figs/img1.png] is //1.15//. This value is remarkably
close to the theoretical value determined in [[2 #papers]], 
which is around [figs/img112.png].

----------------------------------------

===Ordering Step===

The procedure Ordering ([figs/img32.png], [figs/img63.png], [figs/img70.png]) receives 
as input the graph [figs/img32.png] and partitions [figs/img32.png] into the two 
subgraphs [figs/img63.png] and [figs/img70.png], so that [figs/img71.png].

Figure 2 presents a sample graph with 9 vertices
and 8 edges, where the degree of a vertex is shown besides each vertex.
Initially, all vertices with degree 1 are added to a queue [figs/img136.png].
For the example shown in Figure 2(a), [figs/img137.png] after the initialization step.

 | [figs/img138.png]
 | **Figure 2:** Ordering step for a graph with 9 vertices and 8 edges.

Next, we remove one vertex [figs/img139.png] from the queue, decrement its degree and
the degree of the vertices with degree greater than 0 in the adjacent
list of [figs/img139.png], as depicted in Figure 2(b) for [figs/img140.png].
At this point, the adjacencies of [figs/img139.png] with degree 1 are
inserted into the queue, such as vertex 1.
This process is repeated until the queue becomes empty.
All vertices with degree 0 are non-critical vertices and the others are
critical vertices, as depicted in Figure 2(c).
Finally, to determine the vertices in [figs/img141.png] we collect all 
vertices [figs/img142.png] with at least one vertex [figs/img143.png] that 
is in Adj[figs/img144.png] and in [figs/img145.png], as the vertex 8 in Figure 2(c).
 
----------------------------------------

===Searching Step===

In the searching step, the key part is
the //perfect assignment problem//: find [figs/img153.png] such that
the function [figs/img154.png] defined by

 | [figs/img155.png]

is a bijection from [figs/img156.png] to [figs/img157.png] (recall [figs/img158.png]).
We are interested in a labelling [figs/img72.png] of
the vertices of the graph [figs/img59.png] with
the property that if [figs/img11.png] and [figs/img22.png] are keys 
in [figs/img20.png], then [figs/img159.png]; that is, if we associate
to each edge the sum of the labels on its endpoints, then these values
should be all distinct.
Moreover, we require that all the sums [figs/img160.png] ([figs/img18.png])
fall between [figs/img115.png] and [figs/img161.png], and thus we have a bijection
between [figs/img20.png] and [figs/img157.png].

The procedure Searching ([figs/img32.png], [figs/img63.png], [figs/img70.png], [figs/img37.png])
receives as input [figs/img32.png], [figs/img63.png], [figs/img70.png] and finds a 
suitable [figs/img162.png] bit value for each vertex [figs/img74.png], stored in the
array [figs/img37.png].
This step is first performed for the vertices in the
critical subgraph [figs/img63.png] of [figs/img32.png] (the 2-core of [figs/img32.png]) 
and then it is performed for the vertices in [figs/img70.png] (the non-critical subgraph
of [figs/img32.png] that contains the "acyclic part" of [figs/img32.png]).
The reason the assignment of the [figs/img37.png] values is first
performed on the vertices in [figs/img63.png] is to resolve reassignments
as early as possible (such reassignments are consequences of the cycles
in [figs/img63.png] and are depicted hereinafter).

----------------------------------------

====Assignment of Values to Critical Vertices====

The labels [figs/img73.png] ([figs/img142.png])
are assigned in increasing order following a greedy
strategy where the critical vertices [figs/img139.png] are considered one at a time,
according to a breadth-first search on [figs/img63.png].
If a candidate value [figs/img11.png] for [figs/img73.png] is forbidden
because setting [figs/img163.png] would create two edges with the same sum,
we try [figs/img164.png] for [figs/img73.png]. This fact is referred to 
as a //reassignment//.

Let [figs/img165.png] be the set of addresses assigned to edges in [figs/img166.png].
Initially [figs/img167.png].
Let [figs/img11.png] be a candidate value for [figs/img73.png].
Initially [figs/img168.png].
Considering the subgraph [figs/img63.png] in Figure 2(c),
a step by step example of the assignment of values to vertices in [figs/img63.png] is 
presented in Figure 3.
Initially, a vertex [figs/img139.png] is chosen, the assignment [figs/img163.png] is made
and [figs/img11.png] is set to [figs/img164.png].
For example, suppose that vertex [figs/img169.png] in Figure 3(a) is 
chosen, the assignment [figs/img170.png] is made and [figs/img11.png] is set to [figs/img96.png].

 | [figs/img171.png]
 | **Figure 3:** Example of the assignment of values to critical vertices.

In Figure 3(b), following the adjacent list of vertex [figs/img169.png],
the unassigned vertex [figs/img115.png] is reached.
At this point, we collect in the temporary variable [figs/img172.png] all adjacencies 
of vertex [figs/img115.png] that have been assigned an [figs/img11.png] value, 
and [figs/img173.png].
Next, for all [figs/img174.png], we check if [figs/img175.png].
Since [figs/img176.png], then [figs/img177.png] is set 
to [figs/img96.png], [figs/img11.png] is incremented 
by 1 (now [figs/img178.png]) and [figs/img179.png].
Next, vertex [figs/img180.png] is reached, [figs/img181.png] is set 
to [figs/img62.png], [figs/img11.png] is set to [figs/img180.png] and [figs/img182.png].
Next, vertex [figs/img183.png] is reached and [figs/img184.png].
Since [figs/img185.png] and [figs/img186.png], then [figs/img187.png] is 
set to [figs/img180.png], [figs/img11.png] is set to [figs/img183.png] and [figs/img188.png].
Finally, vertex [figs/img189.png] is reached and [figs/img190.png].
Since [figs/img191.png], [figs/img11.png] is incremented by 1 and set to 5, as depicted in 
Figure 3(c).
Since [figs/img192.png], [figs/img11.png] is again incremented by 1 and set to 6, 
as depicted in Figure 3(d).
These two reassignments are indicated by the arrows in Figure 3.
Since [figs/img193.png] and [figs/img194.png], then [figs/img195.png] is set 
to [figs/img196.png] and [figs/img197.png]. This finishes the algorithm.

----------------------------------------

====Assignment of Values to Non-Critical Vertices====

As [figs/img70.png] is acyclic, we can impose the order in which addresses are
associated with edges in [figs/img70.png], making this step simple to solve
by a standard depth first search algorithm.
Therefore, in the assignment of values to vertices in [figs/img70.png] we
benefit from the unused addresses in the gaps left by the assignment of values
to vertices in [figs/img63.png].
For that, we start the depth-first search from the vertices in [figs/img141.png] because 
the [figs/img37.png] values for these critical vertices were already assigned
and cannot be changed.

Considering the subgraph [figs/img70.png] in Figure 2(c),
a step by step example of the assignment of values to vertices in [figs/img70.png] is 
presented in Figure 4.
Figure 4(a) presents the initial state of the algorithm.
The critical vertex 8 is the only one that has non-critical vertices as
adjacent.
In the example presented in Figure 3, the addresses [figs/img198.png] were not used.
So, taking the first unused address [figs/img115.png] and the vertex [figs/img96.png], 
which is reached from the vertex [figs/img169.png], [figs/img199.png] is set 
to [figs/img200.png], as shown in Figure 4(b).
The only vertex that is reached from vertex [figs/img96.png] is vertex [figs/img62.png], so
taking the unused address [figs/img183.png] we set [figs/img201.png] to [figs/img202.png],
as shown in Figure 4(c).
This process is repeated until the UnAssignedAddresses list becomes empty.
 
 | [figs/img203.png]
 | **Figure 4:** Example of the assignment of values to non-critical vertices.
 
----------------------------------------

==The Heuristic==[heuristic]

We now present an heuristic for BMZ algorithm that 
reduces the value of [figs/img1.png] to any given value between //1.15// and //0.93//.
This reduces the space requirement to store the resulting function 
to any given value between [figs/img12.png] words and [figs/img13.png] words.
The heuristic reuses, when possible, the set 
of [figs/img11.png] values that caused reassignments, just before 
trying [figs/img164.png].
Decreasing the value of [figs/img1.png] leads to an increase in the number of 
iterations to generate [figs/img32.png].
For example, for [figs/img244.png] and [figs/img6.png], the analytical expected number 
of iterations are [figs/img245.png] and [figs/img246.png], respectively (see [[2 #papers]] 
for details),
while for [figs/img128.png] the same value is around //2.13//.

----------------------------------------

==Memory Consumption==

Now we detail the memory consumption to generate and to store minimal perfect hash functions
using the BMZ algorithm. The structures responsible for memory consumption are in the 
following:
- Graph:
  + **first**: is a vector that stores //cn// integer numbers, each one representing 
    the first edge (index in the vector edges) in the list of 
    edges of each vertex. 
    The integer numbers are 4 bytes long. Therefore,
    the vector first is stored in //4cn// bytes.
    
  + **edges**: is a vector to represent the edges of the graph. As each edge
    is compounded by a pair of vertices, each entry stores two integer numbers 
    of 4 bytes that represent the vertices. As there are //n// edges, the 
    vector edges is stored in //8n// bytes. 
    
  + **next**: given a vertex [figs/img139.png], we can discover the edges that 
    contain [figs/img139.png] following its list of edges, 
    which starts on first[[figs/img139.png]] and the next
    edges are given by next[...first[[figs/img139.png]]...]. Therefore, the vectors first and next represent 
    the linked lists of edges of each vertex. As there are two vertices for each edge,
    when an edge is iserted in the graph, it must be inserted in the two linked lists 
    of the vertices in its composition. Therefore, there are //2n// entries of integer
    numbers in the vector next, so it is stored in //4*2n = 8n// bytes.
    
  + **critical vertices(critical_nodes vector)**: is a vector of //cn// bits, 
    where each bit indicates if a vertex is critical (1) or non-critical (0). 
    Therefore, the critical and non-critical vertices are represented in //cn/8// bytes.
    
  + **critical edges (used_edges vector)**: is a vector of //n// bits, where each 
    bit indicates if an edge is critical (1) or non-critical (0). Therefore, the 
    critical and non-critical edges are represented in //n/8// bytes. 
    
- Other auxiliary structures 
  + **queue**: is a queue of integer numbers used in the breadth-first search of the
    assignment of values to critical vertices. There is an entry in the queue for 
    each two critical vertices. Let [figs/img110.png] be the expected number of critical 
    vertices. Therefore, the queue is stored in //4*0.5*[figs/img110.png]=2[figs/img110.png]//.
    
  + **visited**: is a vector of //cn// bits, where each bit indicates if the g value of 
    a given vertex was already defined. Therefore, the vector visited is stored
    in //cn/8// bytes.
    
  + **function //g//**: is represented by a vector of //cn// integer numbers.
    As each integer number is 4 bytes long, the function //g// is stored in
    //4cn// bytes. 

    
Thus, the total memory consumption of BMZ algorithm for generating a minimal 
perfect hash function (MPHF) is: //(8.25c + 16.125)n +2[figs/img110.png] + O(1)// bytes.
As the value of constant //c// may be 1.15 and 0.93 we have:
 || //c// |  [figs/img110.png] | Memory consumption to generate a MPHF |
  | 0.93  |  //0.497n//  |         //24.80n + O(1)//             |
  | 1.15  |  //0.401n//  |         //26.42n + O(1)//             |
  
  | **Table 1:** Memory consumption to generate a MPHF using the BMZ algorithm.
  
The values of [figs/img110.png] were calculated using Eq.(1) presented in [[2 #papers]].
    
Now we present the memory consumption to store the resulting function.
We only need to store the //g// function. Thus, we need //4cn// bytes.
Again we have:
 || //c// | Memory consumption to store a MPHF |
  | 0.93  |            //3.72n//               |
  | 1.15  |            //4.60n//               |
  
  | **Table 2:** Memory consumption to store a MPHF generated by the BMZ algorithm.  
----------------------------------------

==Experimental Results==

[CHM x BMZ comparison.html]

----------------------------------------

==Papers==[papers]

+ [F. C. Botelho http://www.dcc.ufmg.br/~fbotelho], D. Menoti, [N. Ziviani http://www.dcc.ufmg.br/~nivio]. [A New algorithm for constructing minimal perfect hash functions papers/bmz_tr004_04.ps], Technical Report TR004/04, Department of Computer Science, Federal University of Minas Gerais, 2004.

+ [F. C. Botelho http://www.dcc.ufmg.br/~fbotelho], Y. Kohayakawa, and [N. Ziviani http://www.dcc.ufmg.br/~nivio]. [A Practical Minimal Perfect Hashing Method papers/wea05.pdf]. //4th International Workshop on efficient and Experimental Algorithms (WEA05),// Springer-Verlag Lecture Notes in Computer Science, vol. 3505, Santorini Island, Greece, May 2005, 488-500.


%!include: ALGORITHMS.t2t

%!include: FOOTER.t2t

%!include(html): ''GOOGLEANALYTICS.t2t''