prev

next

out of 72

View

214Download

0

Embed Size (px)

8/14/2019 Design Pattern for Graph Algorithms

1/72

Design Patterns for the Implementation of Graph Algorithms

Dietmar Kuhl

Technische Universitat Berlin

Berlin, the 19th July 1996

8/14/2019 Design Pattern for Graph Algorithms

2/72

8/14/2019 Design Pattern for Graph Algorithms

3/72

Contents

1 Introduction 2

1.1 Reuse of Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Existing Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Outline of the Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Problem Domain 7

2.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Algorithms on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3 Graph Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 General Approaches 31

3.1 Object-Oriented Software Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2 Design Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.3 The C++ Standard Template Library . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Concepts 36

4.1 The Big Picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2 Abstraction from the Representation . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.2.1 Adjacency Iterator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.2.2 Data Accessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.2.3 Modifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.3 Algorithm Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.3.1 Algorithm Step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.3.2 Initializer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.3.3 Loop Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5 Conclusions 68

1

8/14/2019 Design Pattern for Graph Algorithms

4/72

Chapter 1

Introduction

It is observed in [1] that in software design the wheel is continually reinvented. Thus it is proposedto implement an Algorithmic Tool Kit. This Algorithmic Tool Kit is supposed to include graph

algorithms. Although the aim to create such a tool kit is clearly stated, the paper gives no insight how

this tool kit looks like: The exact approach how to implement such a tool kit is left open.

When writing software there are some important objectives to be achieved: Efficiency, correct-

ness, and reusability are the most common objectives. None of them can be completely ignored and

correctness is even essential. This work will provide a concept how to write reusable but still efficient

implementations of graph algorithms. The topic of provable correctness of the implementation is not

covered although the concepts given should still be applicable where provable correctness is required.

This chapter introduces the reasons for concentrating on reusability. It briefly reviews existing

approaches to implement graph algorithms and gives an overview of the concepts proposed. At the

end of this chapter an outline for the remainder of the work is given.

1.1 Reuse of Implementations

Often, different programs need similar algorithms and data structures to solve some problems. To

avoid implementing an algorithm multiply times, algorithms are implemented once and made avail-

able in a library which can then be used by any program needing an algorithm from the library. Ex-

amples for algorithms found in libraries are sorting algorithms, algorithms to solve linear equations,

or algorithms to compute an optimal solution of a linear program. There are several reasons to reuse

an algorithms rather than reimplementing it:

For most algorithms it takes some expert knowledge to implement them. This already starts withrelatively simple algorithms like sorting algorithms: Although sorting algorithms are very basic

and topic of most beginner courses in computer science, they are sufficient complex to make a

lookup in a book necessary to find out the details. More complex algorithms will require more

knowledge about the domain, e.g. to guarantee numerical correctness or to be as efficient as

possible.

Implementing any algorithm takes time. To get an algorithm implemented takes a programmer

which has to spend some time he/she has to be paid for. Thus, using an existent implementation

of an algorithm may save money if it is easier to use the implementation than reimplementing

the algorithm. For sufficient complex algorithms, e.g. finding an optimal solution of a linear

program, reuse of an existent implementation will always pay off. For simple algorithms it

2

8/14/2019 Design Pattern for Graph Algorithms

5/72

may depend on the interface provided to the implementation: If the implementation uses an

inconvenient or hard to use interface, it may be easier and faster to reimplement the algorithm.

In any case, using the same implementation wherever appropriate amortizes the costs of theimplementation between different projects reducing the costs of the individual projects.

Initial implementations of algorithms are often erroneous. The description of the algorithm,

whether it is found in the literature or is developed by the programmer, may contain errors.

If the description of the algorithm is correct, the translation to a programming language may

introduce errors. Of course, an implementation which is reused may also contain errors but

because the implementation is (hopefully) often used, it is more likely that errors are detected

and can be corrected. An implementation which is only created for one application is more

likely to contain undetected errors as the implementation is not that often applied.

When programming an algorithm for a library, this algorithm is the main concern. To make it

a good candidate for inclusion into a library it is not only tested thoroughly but also optimizedto be as efficient as possible. If some algorithm is implemented just to solve a problem arising

in some bigger context, this implementation is typically not optimized at all and often uses

a simple but not efficient algorithm. Thus, an algorithm taken from a library is often more

efficient than a special implementation solving the problem.

If this were the whole story, clearly every algorithm would be taken from some library. Unfor-

tunately, there are also problems with algorithms taken from libraries. An obvious problem are the

data structures used: The more complex the data structures are to represent the problem, the more

choices are available to design the data structure. To use an algorithm from a library the data structure

of the library has to match the data structure of the application somehow. An approach often taken is

to implement an algorithm for a certain data structure which has to be used to apply this algorithm.In areas where there is a canonical representation, e.g. because the representation is efficiently sup-

ported by the programming language, this approach is sufficient. However, this not the case for the

representation of graphs which can be represented in many different ways with different tradeoffs.

Another problem arising with implementations from libraries is the fact that the algorithm can

normally not be modified: Graph algorithms often extend the behavior of an existent algorithm by

making additional computations in each iteration of an algorithm, by adding another termination con-

dition, or by using a special initialization and so on. This requires that the behavior of an algorithm

can be modified making the design of the implementation hard: Often, additional flexibility is traded

for reduced efficiency e.g. because an additional function has to be called or additional conditions

have to be checked.

Despite these problems it would be highly desirable to have reusable implementations of graphalgorithms available: Most of the algorithms working on graphs are very complex and take a con-

siderable amount of insight into the problem domain to understand and implement them. Thus, it

seems to be logical to search for an approach to the implementation of graph algorithms alleviating

the problems encountered.

1.2 Existing Approaches

This section briefly reviews two existing packages: LEDA and SCOPE. The reason to review these

implementations is to identify common coding practice which hinders or enhances reusability.

3

8/14/2019 Design Pattern for Graph Algorithms

6/72

Library of Efficie