Dependability

Dependable software
Bertrand Meyer, ETH Zurich
Cite as follows: Bertrand Meyer, Dependable

Software, to appear in Dependable Systems:
Software, Computing, Networks, eds. Jrg Kohlas,
Bertrand Meyer, Andr Schiper, Lecture Notes in
Computer Science, Springer-Verlag, 2006.
ABSTRACT
Achieving software reliability takes many complementary techniques, directed
at the process or at the products. This survey summarizes some of the most
fruitful ideas.
1 OVERVIEW
Everyone who uses software or relies on devices or processes that use software
in other words, everyone has a natural interest in guarantees that
programs will perform properly. The following pages provide a review of
techniques to improve software quality.
There are many subcultures of software quality research, often seemingly
sealed off from each other; mentioning process-based approaches such as
CMMI to programming language technologists, or tests to people working on
proofs, can be as incongruous as bringing up Balanchine among baseball fans.
This survey disregards such established cultural fences and instead attempts to
include as many as possible of the relevant areas, on the assumption that
producing good software is hard enough that every little bit counts [60]. As
a result we will encounter techniques of very diverse kinds.
A note of warning to the reader seeking objectivity: I have not shied away
from including references easy to spot to my own work, with the
expectation (if a justification is needed) that it makes the result more lively
than a cold inspection limited to other peoples products and publications.
2 SCOPE AND TERMINOLOGY

The first task is to define some of the fundamental terms. Even the first word
of this articles title, determined by the Hasler Foundations Dependable
Information and Communication Systems project, requires clarification.
Reliability and dependability

In the software engineering literature the more familiar term is not
dependable but reliable, as in software reliability. A check through
general-purpose and technical dictionaries confirms that the two have similar
definitions and are usually translated identically into foreign languages.
DEPENDABLE SOFTWARE
There does exist a definition of dependability [1] from the eponymous

IFIP Working Group 10.4 [39] that treats reliability as only one among
dependability attributes, along with availability, safety, confidentiality,
integrity and maintainability. While possibly applicable to a computing
system as a whole, this classification does not seem right for their software
part, as some attributes such as availability are not properties of the software
per se, others such as confidentiality are included in reliability (through one of
its components, security), and the remaining ones such as maintainability are
of dubious meaning for software, being better covered by other quality factors
such as extendibility and reusability [57].
As a consequence of these observations the present survey interprets
dependability as meaning the same thing, for software, as reliability.
Defining reliability
The term software reliability itself lacks a universally accepted definition.
One could argue for taking it to cover all external quality factors such as ease
of use, efficiency and extendibility, and even internal quality factors such as
modularity. (The distinction, detailed in [57], is that external factors are the
properties, immediate or long-term, that affect companies and people
purchasing and using the software, whereas internal factors are perceptible
only to software developers although in the end they determine the attainment
of external factors.)
It is reasonable to retain a more restricted view in which reliability only
covers three external factors: correctness, robustness and security. This
doesnt imply that others are irrelevant; for example even the most correct,
robust and secure system can hardly be considered dependable if in practice it
takes ages to react to inputs, an efficiency problem. The same goes for ease of
use: many software disasters on record happened with systems that
implemented the right functions but made them available through error-prone
user interfaces. The reasons for limiting ourselves to the three factors listed
are, first, that including all others would turn this discussion into a survey of
essentially the whole of software engineering (see [33]); second, that the
techniques to achieve these three factors, although already very diverse, have
a certain kindred spirit, not shared by those for enhancing efficiency (like
performance optimization techniques), ease of use (like ergonomic design)
and other external and internal factors.
Correctness, robustness, security

For the three factors retained, we may rely on the following definitions:
Correctness is a systems ability to perform according to its specification

in cases of use within that specification.
Robustness is a systems ability to prevent damage in cases of erroneous

use outside of its specification.
Security is a systems ability to prevent damage in cases of hostile use

outside of its specification.
They correspond to levels of increasing departure from the specification. The
specification of any realistic system makes assumptions, explicit or implicit,
about the conditions of its use: a C compilers specification doesnt define a
generated program if the input is payroll data, any more than a payroll
program defines a pay check if the input is a C program; and a buildings
access control software specification cannot define what happens if the
building has burned. By nature, the requirements defined by robustness and
security are different from those of correctness: outside of the specification,
we can no longer talk of performing according to that specification, but only
seek the more modest goal of preventing damage; note that this implies the
ability to detect attempts at erroneous or hostile use.
Security deserves a special mention as in recent years it has assumed a
highly visible place in software concerns. This is a phenomenon to be both
lamented, as it signals the end of a golden age of software development when
we could concentrate on devising the best possible functionality without too
much concern about the worlds nastiness, and at the same time taken to
advantage, since it has finally brought home to corporations the seriousness of
software quality issues, a result that decades of hectoring by advocates of
modern software engineering practices had failed to achieve. One of the most
visible signs of this phenomenon is Bill Gatess edict famously halting all
development in February of 2001 in favor of code reviews for hunting down
security flaws. Many of these flaws, such as the most obnoxious, buffer
overflow, are simply the result of poor software engineering practices. Even if
focusing on security means looking at the symptom rather than the cause,
fixing security implies taking a coherent look at software tools and techniques
and requires, in the end, ensuring reliability as a whole.
Product and process

Any comprehensive discussion of software issues must consider two
complementary aspects: product and process.
The products are the software elements whose reliability we are trying to
assess; the process includes the mechanisms and procedures whereby people
and their organizations build these products.
DEPENDABLE SOFTWARE
The products of software

The products themselves are diverse. In the end the most important one, for
which we may assess correctness, robustness and security, is code. But even
that simple term covers several kinds of product: source code as programmers
see it, machine code as the computer executes it, and any intermediate versions
as exist on modern platforms, such as the bytecode of virtual machines.
Beyond code, we should consider many other products, which in their
own ways are all software: requirements, specifications, design diagrams
and other design documents, test data but also test plans , user
documentation, teaching aids
To realize why it is important in the search for quality to pay attention to
products other than code, it suffices to consider the results of numerous
studies, some already decades old [10], showing the steep progression of the
cost of correcting an error the later it is identified in the lifecycle.
Deficiencies
In trying to ascertain the reliability of a software product or process we must
often like a detective or a fire prevention engineer adopt a negative
mindset and look for sources of violation of reliability properties. The
accepted terminology here distinguishes three levels:
A failure is a malfunction of the software. Note that this term does not
directly apply to products other than executable code.
A fault is a departure of the software product from the properties it should

have satisfied. A failure always comes from a fault, although not
necessarily a fault in the code: it could be in the specification, in the
documentation, or in a non-software product such as the hardware on
which the system runs.
An error is a wrong human decision made during the construction of the

system. Wrong is a subjective term, but for this discussion its clear
what it means: a decision is wrong if it can lead to a fault (which can in
turn cause failures).
In a discussion limited to software reliability, all faults and hence all failures
result from errors, since software is an intellectual product not subject to the
slings and arrows of the physical world.
The more familiar term for error is bug. The upper crust of the software
engineering literature shuns it for its animist connotations. Error has the
benefit of admitting that our mistakes dont creep into our software: we insert
them ourselves. In practice, as may be expected, everyone says bug.
Verification and validation

Even with subjectivity removed from the definition of error, definitions for
the other two levels above remains relative: what constitutes a malfunction
(for the definition of failures) or a departure from desirable properties (for
faults) can only be assessed with respect to some description of the expected
characteristics.
While such reference descriptions exist for some categories of software
product an element of code is relative to a design, the design is relative to
a specification, the specification is relative to an analysis of the requirements
the chain always stops somewhere; for example one cannot in the end certify
that the requirements have no fault, as this would mean assessing them against
some higher-level description, and would only push the problem further to
assessing the value of the description itself. Turtles all the way up.
Even in the absence of another reference (another turtle) against which to
assess a particular product, we can often obtain some evaluation of its quality
by performing internal checks. For example:
A program that does not initialize one of its variables along a particular
path is suspicious, independently of any of its properties vis--vis the
fulfillment of its specification.
A poorly written user manual may not explicitly violate the prescriptions
of another project document, but is problematic all the same.
This observation leads to distinguishing two complementary kinds of
reliability assessment, verification and validation, often combined in the
abbreviation V&V:
Verification is internal assessment of the consistency of the product,

considered just by itself. The last two examples illustrated properties that
are subject to verification: for code; for documentation. Type checking is
another example.
Validation is relative assessment of a product vis--vis another that

defines some of the properties that it should satisfy: code against design,
design against specification, specification against requirements,
documentation against standards, observed practices against company
rules, delivery dates against project milestones, observed defect rates
against defined goals, test suites against coverage metrics.
A popular version of this distinction [10] is that verification is about
ascertaining that the product is doing things right and validation that it is
doing the right thing. It only applies to code, however, since a specification,
a project plan or a test plan do not do anything.
DEPENDABLE SOFTWARE
3 CLASSIFYING APPROACHES
One of the reasons for the diversity of approaches to software quality is the
multiplicity of problems they address. The following table shows a list of
criteria, essentially orthogonal, for classifying them.
Criteria for classifying approaches to software reliability

A priori (build)
A posteriori (assess and correct)
Process
Product
Manual
Tool-supported
Technology-neutral
Technology-specific
Product- and phase-neutral

Static (uses software text)
vs
Product- or phase-specific
Dynamic (requires execution)
Informal
Mathematical
Complete (guarantee)
Partial (some progress)
Free
Commercial
The first distinction is cultural almost as much as it is technical. With a priori

techniques the emphasis is methodological: telling development teams to
apply certain rules to produce a better product. With a posteriori techniques,
the goal is to examine a proposed software product or process element for
possible deficiencies, with the aim of correcting them. While it is natural to
state that the two are complementary rather than contradictory a defense
often used by proponents of a posteriori approaches such as testing when
criticized for accepting software technology as it is rather than helping to
improve it they correspond to different views of the software world, one
hopeful of prevention and the other willing to settle down for cure.
The second distinction corresponds to the two dimensions of software
engineering cited above: are we working on the products, or on the processes
leading to them?
Some approaches are of a methodological nature and just require
applying some practices; we may call them manual, in contrast with
techniques that are tool-supported and hence at least partially automated.
3 CLASSIFYING APPROACHES
An idea can be applicable regardless of technology choices; for example

process-based techniques such as CMMI, discussed below, explicitly stay
away from prescribing specific technologies. At the other extreme, certain
techniques may be applicable only if you accept a certain programming
language, specification method, tool or other technology choice. We may talk
of technology-neutral and technology-specific approaches; this is more a
spectrum of possibilities than a black-and-white distinction, since many
approaches assume a certain class of technologies such as object-oriented
development encompassing many variants.
Some techniques apply to a specific product or phase of the lifecycle:
specification (a specification language), implementation (a static analyzer of
code) They are product-specific, or phase-specific. Others, such as
configuration management tools, apply to many or all product kinds; they are
product-neutral. Product is used here to denote one of the types of outcome
of the software construction process.
For techniques directed at program quality, an important division exists
between dynamic approaches such as testing, which rely on executing the
program, and purely static ones, such as static analysis and program proofs,
which only need to analyze the program text. Here too some nuances exist: a
simulation technique requires execution and hence can be classified as
dynamic even though the execution doesnt use the normal run-time
environment; model-checking is classified as static even though in some
respect it is close to testing.
Some methods are based on mathematical techniques; this is obviously
the case with program proofs and formal specification in general. Many are
more informal.
A technique intended to assess quality properties can give you a complete
guarantee that they are satisfied, or more commonly some partial
reassurance to this effect.
The final distinction is economic: between techniques in the public
domain usable for free, in the ordinary sense of the term and
commercial ones.
DEPENDABLE SOFTWARE
4 PROCESS-BASED APPROACHES
We start with the least technical approaches, emphasizing management
procedures and organizational techniques.
Lifecycle models
One of the defining acts of software engineering was the recognition of the
separate activities involved, in the form of lifecycle models that prescribe a
certain order of tasks (see the figure on the adjacent page). The initial model
is the so-called waterfall [11], still used as a reference for discussions of the
software process although no longer recommended for literal application.
Variants include:
The V model which retains the sequential approach of the waterfall but
divides the process into two parts, the branches of the V; activities
along the first branch are for development, those in the second branch are
for verification and validation, each applied to the results of one of the
steps along the first branch.
The Spiral model [11] which focuses on reducing risk in project

management, in particular the risk caused by the all-or-nothing attitude
of the Waterfall approach. The spiral model suggests isolating subsets of
the systems functionality that are small enough to be implemented
quickly, and when they have been implemented taking advantage of the
experience to proceed to other parts of the system. The idea is connected
with the notion of rapid prototyping.
The Rational Unified Process, distinguishing four phases, inception,

elaboration, construction and transition, with a spiral-like iterative style
of development and a set of recommended best practices such as
configuration management.
The Cluster model [51] [57], emphasizing a different form of

incrementality building a system by layers, from the most
fundamental to the most user-oriented and a seamless process treating
successive activities, from analysis to design, implementation and
maintenance, as a continuum. This model also introduces, as part of the
individual lifecycle of every cluster, a generalization step to prepare for
future reuse of some of the developed elements.
The figure shows pictorial representations of some of these models.
Waterfall
Cluster
V-shaped
Spiral (from [11])
Lifecycle models, illustrated
10
DEPENDABLE SOFTWARE
Whatever their effect on how people actually develop software, the

contribution of lifecycle models has been a classification and definition of the
activities involved in software development, even when these activities are not
executed as phases in the precise order mandated by, for example, the
waterfall model. Software quality benefits in particular from:
A distinction between requirements, the recording of user requirements,

and specification, their translation into a systematic form suitable for
software development, where rigor and precision are essential.
Recognition of the importance of Verification and Validation tasks.
Recognition of post-delivery activities such as maintenance, although

they still do not occupy a visible enough place. Many software troubles
result from evolutions posterior to the initial release.
In the Cluster model, the presence, for each cluster, of the generalization
task to prepare for reuse.
Also in the Cluster model, the use of a seamless and reversible approach
which unifies the methods, tools, techniques and notations that help
throughout the software process, rather than exaggerate them. (The textbook
counter-example here is the use of UML for analysis and design [56].)
The growing emphasis on incrementality in the development process,

even if this concept is understood differently in, for example, the spiral,
cluster and RUP models.
Organizational standards
Another process-related set of developments has had a major effect, largely
beneficial, on some segments of the industry. In the early 1990s the US
Department of Defense, concerned with the need to assess its suppliers
software capabilities and to establish consistent standards, entrusted the
Software Engineering Institute with the task of developing a Capability
Maturity Model, whose current incarnation, CMMI [74] (the I is for
Integration) provides a collection of standards applicable to various
disciplines, rather than a single model for software. Largely independently, the
International Standard Organization has produced a set of software-oriented
variants of its 9000-series quality standards, which share a number of
properties with CMMI. The present discussion is based on CMMI.
Beyond its original target community, CMM and CMMI have been the
catalyst for one of the major phenomena of the IT industry starting in the midnineties: the development of offshore software production, especially in India
[63]. CMMI qualification provides suppliers of outsourcing development
services with quality standards and the associated possibility of independent
certification, without which customers would not be have known how to trust
distant, initially unknown contractors.
CMMI is (in the earlier classification) product-neutral, phase-neutral and
technology-neutral. In its application to software it is intended only to
determine how well an organization controls its development process by
defining and documenting it, recording and assessing how it is applied in
practice, and working to improve it. It doesnt prescribe what the process
should be, only how much you are on top of it. You could presumably be
developing in PL/I on IBM 370 and get CMMI qualification.
CMMI assesses both the capability level of individual process areas in
(such as software) in an organization, and the maturity of an organization as a
whole. It distinguishes five levels of increasing maturity:
Performed: projects happen and results get produced, but there is little
control and no reproducibility; the process is essentially reactive.
Managed: processes are clearly defined for individual projects, but not
for the organization as a whole. They remain largely reactive.
Defined: proactive process defined for the organization.
Quantitatively managed: the control mechanisms do not limit themselves

to qualitative techniques, but add well-defined numerical measurements.
Optimizing: the mechanisms for controlling processes are sufficiently

well established that the focus can shift on improving the organization
and its processes.
Through their emphasis on the process and its repeatability, CMMI and ISO
standards help improve the quality of software development. One may expect
such improvements of the process to have a positive effect on the resulting
products as well; but they are only part of the solution. After a software error
one module of the software was expecting measures in the metric system,
another was providing them in English units was identified as the cause of
the failure of the NASA Mars Orbiter Vehicle mission [82], an engineer from
the project noted that the organization was heavily into ISO and other process
standards. Process models and process-focused practices are not a substitute
for using the best technological solutions. Tailored versions of CMMI that
would not shy away from integrating specific technologies such as object
technology could be extremely useful. In the meantime, the technologyneutral requirements of CMMI can be applied by organizations to get a better
hold on their software processes.
11
12
DEPENDABLE SOFTWARE
Extreme programming
The Extreme Programming movement [6] is a reaction against precisely the
kinds of lifecycle models and process-oriented approaches just reviewed. XP
(as it is also called) emphasizes instead the primacy of code. Some of the
principal ideas include:
Short release cycles to get frequent feedback.
Pair programming (two people at a keyboard and terminal).
Test-driven development.
A general distrust of specification and design: testing is the preferred

guide of development.
Emphasis on programmers welfare.
Some of these practices are clearly beneficial to quality but were developed
prior to XP, in particular short release cycles (Microsofts daily build as
described in 1995 by Cusumano and Shelby [19], see also [54]) and the use of
frequent testing as part of development (see e.g. quality first [55]). Those
really specific to XP are of limited interest (while sometimes a good practice,
pair programming cannot be imposed indiscriminately, both because it doesnt
work for some people and because those who find it useful may not find it
useful all the time) or, in the case of tests viewed as a replacement for
specifications, downright detrimental. See [75] and [64] for critiques of
the approach.
Code inspections
A long-established quality practice is the inspection, also known as review: a
session designed to examine a certain software element with the aim of finding
flaws. The most common form is code inspection, but the process can be
applied to any kind of software engineering product. Rules include:
Small meeting: at most 8 people or so, including the developer of the

element under review.
The elements under review and any supporting documents must be

circulated in advance; the participants should have read them and
identified possible criticisms before the meeting. The allotted time should
be bounded, for example 2 or 3 hours.
The meeting must have a moderator to guide discussions and a secretary

to record results.
The moderator should not be the developers manager. The intent is to

evaluate products, not people.
The sole goal is to identify deficiencies and confirm that they are indeed
deficiencies; correction is not part of the process and should not be
attempted during the meeting.
Code inspections can help avoid errors, but to assess their usefulness one must
compare the costs with those of running automated tools that can catch some
of the same problems without human intervention; static analyzers, discussed
below, are an example.
Some companies have institutionalized the rule that no developer may
check in code (integrate it into the repository for a current or future product)
without approval by one other developer, a limited form of code inspection
that has a clearly beneficial effect by forcing the original developer to convince
at least one other team member of the suitability of the contribution.
Open-source processes
A generalization of the idea of code inspection is the frequent assertion, by
members of the open-source community, that the open-source process
dramatically improves quality by enabling many people to take a critical look
at the software text; some have gone so far as to state that given enough eyes,
all bugs are shallow [73].
As with many of the other techniques reviewed, we may see in this idea
a beneficial contribution, but not a panacea. John Viega gives [78] the example
of a widely used security program in which in the past two years, several very
subtle buffer overflow problems have been found Almost all had been in the
code for years, even though it had been examined many times by both hackers
and security auditors One tool was able to identify one of the problems as
potentially exploitable, but researchers examined the code thoroughly and
came to the conclusion that there was no way the problem could be exploited.
(The last observation is anecdotal evidence for the above observation that tools
such as static analyzers are potentially superior to human analysis.)
While is no evidence that open-source software as a whole is better (or
worse) than commercial software, and no absolute rule should be expected if
only because of the wide variety of products and processes on both sides, it is
clear that more eyes potentially see more bugs.
13
14
DEPENDABLE SOFTWARE
Requirements engineering
In areas such as embedded systems, many serious software failures have been
traced [45] to inadequate requirements rather than to deficiencies introduced
in later phases. Systematic techniques for requirements analysis are available
[76] [40] to improve this critical task of collecting customer wishes and
translating them into a form that can serve as a basis for a software project.
Design patterns
A process-related advance that has had a strong beneficial effect on software
development is the emergence of design patterns [32]. A pattern is an
architectural scheme that has been recognized as fruitful through frequent use
in applications, and for which a precise description exists according to a
standard format. Patterns provide a common vocabulary to developers, hence
simplifying design discussions, and enable them to benefit from the collective
wisdom of their predecessors.
A (minority) view of patterns [62] [65] understands them as a first step
towards the technique discussed next, reusable components. Patterns, in this
interpretation, suffer from the limitation that each developer must manually
insert the corresponding solutions into the architecture of every applicable
system. If instead it is possible to turn the pattern into a reusable component,
developers can directly reuse the corresponding solution through an API
(Abstract Program Interface). The observation here is that it is better to reuse
than to redo. Investigations [65] suggest that with the help of appropriate
programming language constructs up to two thirds of common design patterns
can be thus componentized.
Trusted components
Quality improvement techniques, whether they emphasize the process or the
product, are only as good as their actual application by programmers. The
magnitude of the necessary education effort is enough to temper any hope of
major short-term improvements, especially given that many programmers
have not had the benefit of a formal computer science education to start with.
Another practical impediment to continued quality improvement comes

from market forces. The short-term commercial interest of a company is
generally to release software that is good enough [83]: software that has
barely passed the threshold under which the market would reject it because of
bad quality; not excellent software. The extra time and expense to go from the
first to the second stage may mean, for the company, losing the market to a less
scrupulous competitor, and possibly going out of business. For the industry as
a whole, software quality has indeed improved regularly over time but tends
to peak below the optimum.
An approach that can overcome these obstacles is increased reliance on
reusable components, providing pre-built solutions to problems that arise in
many different applications, either regardless of the technical domain (generalpurpose component libraries) or in particular fields (specialized libraries).
Components have already changed the nature of software development by
providing conveniently packaged implementations, accessible through
abstract interfaces, of common aspects such as graphical user interfaces,
database manipulation, basic numerical algorithms, fundamental data
structures and others, thereby elevating the level at which programmers write
their applications. When the components themselves are of good quality, such
reuse has highly beneficial effects since developers can direct their efforts to
the quality of the application-specific part of their programs.
Examining more closely the relationship of components to quality
actually highlights two separate effects: it is comforting to know that the
quality of a system will benefit from the quality of its components; but we
must note that reuse magnifies the bad as well as the good: imperfections can
be even more damaging in components than in one-of-a-kind developments,
since they affect every application that relies on a component.
The notion of trusted component [58] [61] follows from this analysis that
one of the most pressing and promising tasks for improving software quality
is the industrial production of reusable components equipped with a guarantee
of quality. Producing such trusted components may involve most of the
techniques discussed elsewhere in this article. For some of the more difficult
ones, such as program proving, application to components may be the best
way to justify the cost and effort and recoup the investment thanks to the
scaling effect of component reuse: once a component has reached the level of
quality at which it can really be trusted, it will benefit every application that
relies on it.
15
16
DEPENDABLE SOFTWARE
5 TOOLS AND ENVIRONMENTS

Transitioning now to product-oriented solutions, we examine some of the
progress in tools available to software developers to the extent that it is
relevant for software quality.
Configuration management
Configuration management is a both practice (for the software developer) and
a service (from the supporting tools), so it could in principle be classified
under process as well as under product. It belongs more properly to the
latter category since its tools that make configuration management realistic;
applied as a pure organizational practice without good tool support, it quickly
becomes tedious and ceases being applied.
Configuration management may be defined as the systematic collecting
and registering of project elements, including in particular the ability to:
Register a new version of any project element.
Retrieve any previously registered version of any project element.
Register dependencies, both between project elements and between

registered versions of project elements (e.g. A relies on B, and version 10
of A requires version 7, 8 or 9 of B).
Construct composite products from their constituents for example,

build an executable version of a program from its modules or
reconstruct earlier versions, in accordance with registered dependencies.
A significant number of software disasters on record followed from

configuration management errors, typically due to reintroducing an obsolete
version of a module when compiling a new release of a program, or using an
obsolete version of some data file. Excuses no longer exist for such errors, as
acceptable configuration management tools, both commercial and opensource, are widely available. These tools, while still far from what one could
hope for, have made configuration management one of the most important
practices of modern software development.
Source code is not the only beneficiary of configuration management.
Any product that evolves, has dependencies on other elements and may need
restoring to an earlier state should be considered for inclusion in the
configuration management repository. Besides code this may include project
plans, specification and design documents, user manuals, training documents
such as PowerPoint slides, test data files.
5 TOOLS AND ENVIRONMENTS
Metrics and models

If we believe Lord Kelvins (approximate) maxim that all serious study is
quantitative, then software and software development should be susceptible to
measurement, tempered of course by Einsteins equally famous quote that not
everything measurable is worth measuring. A few software properties, process
or product, are at the same time measurable, worth measuring and relevant to
software reliability.
On the process side, cost in its various dimensions is a prime concern.
While it is important to record costs, if only for CMMI-style traceability, what
most project managers want at a particular time is a model to estimate the cost
of a future project or of the remainder of a current project. Such models do exist
and can be useful, at least if the development process is stable and the project is
comparable to previous ones: then by estimating a number of project parameters
and relying on historical data for comparison one can predict costs essentially,
person-months within reasonable average accuracy. A well-known cost
model, for which free and commercial tools are available, is COCOMO II [12].
During the development of a system, faults will be reported. In principle
they shouldnt be comparable to the faults of a material product, since software
is an intellectual product and doesnt erode, wear out or collapse under attack
from the weather. In practice, however, statistical analysis shows that faults in
large projects can follow patterns that resemble those of hardware systems and
are susceptible to similar statistical prediction techniques. That such patterns can
exist is in fact consistent with intuition: if the tests on the last five builds of a
product under development have each uncovered one hundred new bugs each, it
is unlikely that the next iteration will have zero bugs, or a thousand. Software
reliability engineering [69][46] elaborates on these ideas to develop models for
assessing and predicting failures, faults and errors. As with cost models, a
requirement for meaningful predictions is the ability to rely on historical data for
calibration. Reliability models are not widely known, but could help software
projects understand, predict and manage anomalies better.
More generally, numerous metrics have been proposed to provide
quantitative assessments of software properties. Measures of complexity, for
example, include: source lines of code (SLOC), the most primitive, but useful
all the same; function points [25], which count the number of elementary
mechanisms implemented by the software; measures of the complexity of the
control graph, such as cyclomatic complexity [48][49]; and measures
specifically adapted to object-oriented software [35][59]. The EiffelStudio
environment [30] makes it possible to compute many metrics applied to a project
under development, including measures regarding the use of contracts (section
17
18
DEPENDABLE SOFTWARE
8), and to compare them with values on record. While not necessarily meaningful
in isolation, such measures elements are a useful control tool for the manager;
they are in line with the CMMIs insistence that an organization can only reach
the higher levels of process maturity (4 and 5) by moving from the qualitative to
the quantitative, and should be part of the data collected for such an effort.
Static analyzers
Static analyzers are another important category of tools, increasingly
integrated in development environments, whose purpose is to examine the
software text for deficiencies. They lie somewhere between type checkers
(themselves integrated in compilers) and full program provers, and will be
studied below (page 26) after the discussion of proofs.
Integrated development environments

Beyond individual tools the evolution of software development has led to the
widespread of integrated tool suites known as IDEs for Integrated (originally:
Interactive) Development Environments. Among the best known are
Microsofts Visual Studio [66] and IBMs Eclipse [27]; EiffelStudio [30] is
another example. These environments, equipped with increasingly
sophisticated graphical user interfaces, provide under a single roof a whole
battery of mechanisms to write software (editors), manage its evolution
(configuration management), compile it (compilers, interpreters, optimizers),
examine it effectively (browsers), run it and elucidate the sources of faults
(debuggers, testers), analyze it for possible inconsistencies and errors (static
analysis), generate code from design and analysis diagrams or the other way
around (diagramming, Computer-Aided Software Engineering or CASE,
reverse engineering), change architecture in a safe way through toolcontrolled transformations (refactoring), perform measurements as noted
above (metric tools), and other tasks.
This is one of the most active areas in software engineering; programmers,
for whom IDEs are the basic daily tools, are directly interested in their quality,
so that open-source projects such as Eclipse and EiffelStudio benefit from
active community participation. The effect of these advanced frameworks on
software reliability, while diffuse, is undeniable, as their increasing cleverness
supports quality in several ways: finding bugs through static and dynamic
techniques; avoiding new bugs through mechanisms such as refactoring;
generating some of the code without manual intervention; and, more generally,
providing a level of comfort that frees programmers from distractions and lets
them apply their best skills to the hardest issues of software construction.
6 PROGRAMMING LANGUAGES
6 PROGRAMMING LANGUAGES
The evolution of programming languages plays its part in the search for more
reliable software. High-level languages contribute both positively, by
providing higher levels of expression through advanced constructs freeing the
programmer (in the same spirit as modern IDEs) from mundane, repetitive or
irrelevant tasks, and negatively, by ruling out certain potentially unsafe
constructs and, as a result, eradicate entire classes of bugs at the source.
The realization that programming language constructs could exert a
major influence on software quality both through what they offer and what
they forbid dates back to structured programming [22] [20] which, in the early
seventies, led to rejecting the goto as a control structure in favor of more
expressive constructs sequence, conditional, loop, recursion. The next
major step was object-oriented programming, introducing a full new set of
abstractions, in particular the notion of class, providing decomposition based
on object types rather than individual operations, and techniques of
inheritance and genericity.
In both cases the benefit comes largely from being able to reason less
operationally about software. A software text represents many possible
executions, so many in fact that it is hard to understand the program and
hence to get it right by thinking in terms of what happens at execution [22].
Both structured and object-oriented techniques make it possible to limit such
operational thinking and instead understand the abstract properties of future
run-time behaviors by applying the usual rules of logical reasoning.
In drawing the list of programming languages most important
contributions to quality, we must indeed put at the top all the mechanisms that
have to do with structure. With ever larger programs addressing ever more
ambitious goals, the production and maintenance of reliable software requires
safe and powerful modular decomposition facilities. Particularly noteworthy are:
As pointed out, the class mechanism, which provides a general basis for
stable modules with a clear role in the overall architecture.
Techniques for information hiding, which protect modules against

details of other modules, and permit independent evolution of the various
parts of a system.
Inheritance, allowing the classification and systematic organization of

classes into structured collections, especially with multiple inheritance.
Genericity, allowing the construction of type-parameterized modules.
19
20
DEPENDABLE SOFTWARE
Another benefit of modern languages is static typing which requires

programmers to declare types for all the variables and other entities in their
programs, then takes advantage of this information to detect possible
inconsistencies in their use and reject programs, at compilation time, until all
types fit. Static typing is particularly interesting in object-oriented languages
since inheritance supports a flexible type system in which types can be
compatible even if they are not identical, as long as one describes a
specialization of the other.
Another key advance is garbage collection, which frees programmers
from having to worry about the details of memory management and removes
an entire class of errors such as attempts to access a previously freed
memory cell which can otherwise be particularly hard to detect and to
correct, in particular because the resulting failures are often intermittent rather
than deterministic. Strictly speaking, garbage collection is a property of the
language implementation, but its the language definition that makes it
possible, as with modern object-oriented languages, or not, as in languages
such as C that permit arbitrary pointer arithmetic and type conversions.
Exception handling, as present in modern programming languages,
helps improve software robustness by allowing developers to include recovery
code for run-time faults that would otherwise be fatal, such as arithmetic
overflow or running out of memory.
A mechanism that is equally far-reaching in its abstraction benefits is the
closure, delegate or agent [62]. Such constructs wrap operations in
objects that can then be passed around anonymously across modules of a
system, making it possible to treat routines as first-class values. They
drastically simplify certain kinds of software such as numerical applications,
GUI programming and other event-driven (or publish-subscribe) schemes.
The application of programming language techniques to improving
software quality is limited by the continued reliance of significant parts of the
software industry on older languages. In particular:
Operating systems and low-level system-related tend to be written in C,

which retains its attractions for such applications in spite of widely
known deficiencies, such as the possibility of buffer overflow.
The embedded and mission-critical community sometimes prefers to use

low-level languages, including assembly, for fear of the risks potentially
introduced by compilers and other supporting tools.
The Verifying Compiler Grand Challenge [38] [77] is an attempt to support

the development of tools that even with such programming languages
will guarantee, during the process of compiling and thanks to techniques
described in the following sections, the reliability of the programs they process.
7 STATIC VERIFICATION TECHNIQUES

Static techniques work solely from the analysis of the software text: unlike
dynamic techniques such as tests they do not require any execution to verify
software or report errors.
Proofs
Perhaps the principal difference between mathematics and engineering is that
only mathematics allows providing absolute guarantees. Given the proper
axioms, I can assert with total confidence that two plus two equals four. But if
I want to drive to Berne the best assurance I can get that my car will not break
down is a probability. I know its higher than if I just drive it to the suburbs,
and lower than if my goal were Prague, Alma-Ata, Peking or Bombay; I can
make it higher by buying a new, better car; but it will never be one. Even with
the highest attention to quality and maintenance, physical products will
occasionally fail.
Under appropriate assumptions, a program is like a mathematical
proposition rather than a material device: any general property of the program
stating that all executions of the program will achieve a certain goal, or that
at least one possible execution will is either true or false, and whether it is
true or not is entirely determined by the text of the program, at least if we
assume correct functioning of the hardware and of other software elements
needed to carry out program execution (compiler, run-time system, operating
system). Another way of expressing this observation is that a programming
language is similar to a mathematical theory, in which certain propositions are
true and others false, as determined by the axioms and inference rules.
In principle, then, it should be possible to prove or disprove properties of
programs, in particular correctness, robustness and security properties, using
the same rigorous techniques as in the proofs of any mathematical theorem.
This assumes overcoming a number of technical difficulties:
Programming languages are generally not defined as mathematical

theories but through natural-language documents possessing a varying
degree of precision. To make formal reasoning possible requires
describing them in mathematical form; this is known as providing a
mathematical semantics (or formal semantics) to a programming
language and is a huge task, especially when it comes to modeling
advanced mechanisms such as exception handling and concurrency, as
well as the details of computer arithmetic since the computers view of
integers and reals strays from their standard mathematical properties.
21
22
DEPENDABLE SOFTWARE
The theorems to be proved involve specific properties of programs, such

as the value of a certain variable not exceeding a certain threshold at a
certain state of the execution. Any proof process requires the ability to
express such properties; this means extending the programming language
with boolean-valued expressions, called assertions. Common languages
other than Eiffel do not include an assertion mechanism; this means that
programmers will have to resort to special extensions such as JML for
Java [43] (see also Spec#, an extension of the C# language [5]) and
annotate programs with the appropriate assertions. Some tools such as
Daikon help in this process by extracting tentative assertions from the
program itself [31].
In practice the softwares actual operation depends, as noted, on those of

a supporting hardware and software environment; proofs of the software
must be complemented by guarantees about that environment.
Not all properties lend themselves to easy enunciation. In particular,

non-functional properties such as performance (response time,
bandwidth, memory occupation) are hard to model.
More generally, a proof is only as useful as the program properties being

proven. What is being proved is not the perfection of the program in any
absolute sense, nor even its quality, but only that it satisfies the assertions
stated. It is never possible to know that all properties of interest have been
included. This is not just a theoretical problem: security attacks often take
advantage of auxiliary aspects of the programs behavior, which its
design and verification did not take into account.
Even if the language, the context and the properties of interest are fully
specified semantically and the properties relevant, the proof process
remains a challenge. It cannot in any case be performed manually, since
even the proof of a few properties of a moderately sized programs quickly
reaches into the thousands of proof steps. Fully automated proofs are, on
the other hand, generally not possible. Despite considerable advances in
computer-assisted proof technology (for programs as well as other
applications) significant proofs still require considerable user interaction
and expert knowledge.
Of course the effort may well be worthwhile, especially in two cases: lifecritical systems in transportation and defense to which, indeed, much proof
work has been directed; and reusable components, for which the effort is
justified as explained in the discussion of Trusted Components above by
the scaling-up effect of reuse.
Here are some of the basic ideas about how proofs work. A typical
program element to prove would be, in Eiffel notation
decrement
-- Decrease counter by one.
require
counter > 0
do
counter := counter 1
ensure
counter = old counter 1
counter >= 0
end
This has a program body, the do clause, and two assertions, a precondition
introduced by require and a postcondition introduced by ensure and
consisting of two subclauses implicitly connected by an and. Assertions are
essentially boolean expressions of the language with the possibility, in a
postcondition, of using the old notation to refer to values on entry: here the
first subclause of the postcondition states that the value of counter will have
been decreased by one after execution of the do clause.
Program proofs deal with such annotated programs, also called
contracted programs (see section 8 below). The annotations remind us that
proofs and other software quality assurance technique can never give us
absolute guarantees of quality: we can never say that a program is correct,
only assess it whether through rigorous techniques like proofs or using
more partial ones such as those reviewed next relatively to explicitly stated
properties, expressed here through assertions integrated in the program text.
From a programmers viewpoint the above extract is simply the text of a
routine to be executed, with some extra annotations, the precondition and
postcondition, expressing properties to be satisfied before and after. But for
proof purposes this text is a theorem, asserting that whenever the body (the do
clause with its assignment instruction) is executed with the precondition
satisfied it will terminate in such a way that the postcondition is satisfied.
This theorem appears to hold trivially but even before addressing the
concern noted above that computer integers are not quite the same as
mathematical integers proving it requires the proper mathematical
framework. The basic rule of axiomatic semantics (or Hoare semantics [37])
covering such cases is the assignment axiom, which for any variable x and
expression e states that the following holds
23
24
DEPENDABLE SOFTWARE
require Q (e) do x := e ensure Q (x)

where Q (x) is an assertion which may depend on x; then Q (e) is the same
assertion with every mention of x replaced by e, except for occurrences of
old x which must be replaced by x.
This very general axiom captures the properties of assignment (in the
absence of side effect in the evaluation of e); its remarkable feature is that it is
applicable even if the source expression e contains occurrences of the target
variable x, as in the example (where x is counter).
We may indeed apply the axiom to prove the examples correctness. Let
Q1 (x) be x = old x 1, corresponding to the first subclause of the
postcondition, and Q2 (x) be x >= 0. Applying the rule to Q1 (counter), we
replace counter by counter + 1 and old counter by counter; this gives
counter 1 = counter 1, which trivially holds. Applying now the same
transformations to Q2 (counter), we get counter 1 >= 0, which is equivalent
to the precondition counter > 0. This proves the correctness of our little
assertion-equipped example.
From there the theory moves to more complex constructions. An
inference rule states that if you have proved
require P do Instruction_1 ensure Q
and
require Q do Instruction_2 ensure R
(note the postcondition of the first part matching the precondition of the
second part) you are entitled to deduce
require P do Instruction_1 ; Instruction_2 ensure Rt
and so on for more instructions. A rule in the same style enables you to deduce
properties of if c then I1 else I2 end from properties of I1 and I2. More
advanced is the case of loops: to prove the properties of
from
Initialization
until
Exit
loop
Body
end
you need, in this general approach, to introduce a new assertion called the loop
invariant and an integer expression called the loop variant. The invariant is
a weakened form of the desired postcondition, which serves as approximation
of the final goal; for example if the goal is to compute the maximum of a set
of values, the invariant will be Result is the maximum of the values processed
so far. The advantage of the invariant is that it is possible both to:
Ensure the invariant through initialization (the from clause in the above
notation); in the example the invariant will be trivially true if we start with
just one value and set Result to that value.
Preserve the invariant through one iteration of the loop body (the loop
clause); in the example it suffices to extend the set of processed values by
one element v and execute if v > Result then Result := v end.
If indeed a loop possesses such an invariant and its execution terminates, then
on exit the invariant will still hold (since it was ensured by the initialization
and preserved by all the loop iterations), together with the Exit condition. The
combination of these two assertions gives the postcondition of the loop. Seen
the other way around, if we started from a desired postcondition and weakened
it to get an invariant, we will obtain a correct program. In the example, if the
exit condition states that we have processed all values of interest, combining
this property with the invariant Result is the maximum of the values
processed so far tells us that Result is the maximum of all values.
Such reasoning is only interesting if the loop execution actually
terminates; this is where the loop variant comes in. It is an integer expression
which must have a non-negative value after the Initialization and decrease,
while remaining non-negative, whenever the Body is executed with the Exit
condition not satisfied. The existence of such an expression is enough to
guarantee termination since a non-negative integer value cannot decrease
forever. In the example a variant is N i where N is the total number of values
being considered for the maximum (the proof assumes a finite set) and i the
number of values processed.
Axioms and inference rules similarly exist for other constructs of
programming languages, becoming, as noted, more intricate as one moves on
to more advanced mechanisms.
25
26
DEPENDABLE SOFTWARE
For concurrent, reactive and real-time systems, boolean assertions of the

kind illustrated above may not be sufficient; it is often convenient to rely on
properties of temporal logic [47], which given a set of successive
observations of a programs execution, can express, for a boolean property Q:
forever Q: from now on, Q will always hold.
eventually Q: at some point in the future (where future includes now),

Q will hold.
P until Q: Q will hold at some point in the future, and until then P will hold.
Regardless of the kind of programs and properties being targeted, there are
two approaches to producing program proofs. The analytic method takes
programs as they exist, then after equipping them with assertions, either
manually or with some automated aid as noted above, attempts the proof. The
constructive method [24] [2] [68] integrates the proof process in the software
construction process, often using successive refinements to go from
specification to implementation through a sequence of transformations, each
proved to preserve correctness, and integrating more practical constraints at
every step.
Proof technology has had some notable successes, including in industrial
systems (and in hardware design), but until recently has remained beyond the
reach of most software projects.
Static analysis
If hoping for a proof covering all the correctness, reliability and security
properties of potential interest is often too ambitious, the problem becomes
more approachable if we settle for a subset of these properties a subset that
may be very partial but very interesting. For example being able to determine
that no buffer overflow can ever arise in a certain program in other words,
to provide a firm guarantee, through analysis of the program text, that every
index used at run time to access an item in an array or a character in a string
will be within the defined bounds is of great practical value since this rules
out a whole class of security attacks.
Static analysis is the tool-supported analysis of software texts for the
purpose of assessing specific quality properties. Being static, it requires no
execution and hence can in principle be applied to software products other
than code. Proofs are a special case, the most far-reaching, but other static
analysis techniques are available.
At the other extreme, a well-established form of elementary static

analysis is type checking, which benefits programs written in a statically typed
programming language. Type checking, usually performed by the compiler
rather than by a separate tool, ascertains the type consistency of assignments,
routine calls and expressions, and rejects any program that contains a
type incompatibility.
More generally, techniques usually characterized as static analysis lie
somewhere between such basic compiler checks and full program proofs.
Violations that can typically be detected by static analysis include:
Variables that, on some control paths, would be accessed before being

initialized (in languages such as C that do not guarantee initialization).
Improper array and string access (buffer overflow).
Memory properties: attempt to access a freed location, double freeing,

memory leak
Pointer management (again in low-level languages such as C): attempts

to follow void or otherwise invalid pointers.
Concurrency control: deadlocks, data races.
Miscellaneous: certain cases of arithmetic overflow or underflow,

changes to supposedly constant strings
Static analysis tools such as PREfix [72] have been regularly applied for
several years to new versions of the Windows code base and have avoided
many potential errors.
One of the issues of static analysis is the occurrence of false alarms:
inconsistency reports that, on inspection, do not reveal any actual error. This
was the weak point of older static analyzers, such as the widely known Lint tool
which complements the type checking of C compilers: for a large program they
can easily swamp their users under thousand of messages, most of them
spurious, but requiring a manual walkthrough to sort out the good from the bad.
(In the search for errors, of course, the good is what otherwise would be
considered the bad: evidence of wrongdoing.) Progress in static analysis has
been successful in considerably reducing the occurrence of false alarms.
The popularity of static analysis is growing; the current trend is to extend
the reach of static analysis tools ever further towards program proofs. Two
examples are:
27
28
DEPENDABLE SOFTWARE
Techniques of abstract interpretation [18] with the supporting ASTRE

tool [9], which has been used to prove the absence of run-time errors in
the primary flight control software, written in C, for the Airbus A340 flyby-wire system.
ESC-Java [21] and, more recently, the Boogie analyzer [4] make program
proving less obtrusive by incrementally extending the kind of diagnostics
with which programmers are familiar, for example type errors, to more
advanced checks such as the impossibility to guarantee that an invariant
is preserved.
Model checking
The model checking approach to verification [36] [17] [3] is static, like proofs
and static analysis, but provides a natural link to the dynamic techniques
(testing) studied below. The inherent limitation of tests is that they can never
be exhaustive; for any significant system in fact, even for toy examples
the number of possible cases skyrockets into the combinatorial stratosphere,
where the orders of magnitude invite lyrical comparisons with the number of
particles in the universe.
The useful measure is the number of possible states of a program. The
notion of state was implicit in the earlier discussion of assertions. A state is
simply a snapshot of the program execution, as could be observed, if we stop
that execution, by looking up the contents of the programs memory, or more
realistically by using the debugger to examine the values of the programs
variables. Indeed it is the combination of all the variables values that
determines the state. With every 64-bit integer variable potentially having 264
values, it is not surprising that the estimates quickly go galactic.
Model checking attempts exhaustive analysis of program states anyway
by performing predicate abstraction. The idea is to simplify the program by
replacing all expressions by boolean expressions (predicates), with only two
possible values, so that the size of the state space decreases dramatically; it
will still be large, but the power of modern computers, together with smart
algorithms, can make its exploration tractable. Then to determine that a
desired property holds for example, a security property such as the absence
of buffer overflows, or a timing property such as the absence of deadlock it
suffices to evaluate the corresponding assertion in all of the abstract states and,
if a violation of that assertion (or counter-example) is found, to check that it
also arises in the original program.
For example, predicate abstraction will reduce a conditional instruction

if a > b then... to if p then..., where p is a boolean. This immediately cuts down
the number of cases from 2128 to 2. The drawback is that the resulting program
is only a caricature of the original; it loses the relation of p to other predicates
involving a and b. But it has an interesting property: if the original violates the
assertion, then the abstracted version also does. So the next task is to look for
any such violation in the abstracted version. This may be possible through
exhaustive examination of its reduced state space, and if so is guaranteed to
find any violation in the original program, but even so is not the end of the story,
since the reverse proposition does not hold: a counter-example in the abstracted
program does not necessarily signal a counter-example in the original. It could
result from the artificial merging of several cases, for example if it occurs on
a path impossible in an execution of the original program obtained by
selecting both p and q as true where q is the abstraction of b > a + 1. Then
examining the state space of the abstracted program will either:
Not find any violations, in which case it proves there was none in the
original program.
Report violations, each of which might be an error in the original or

simply a false alarm generated by the abstraction process.
So the remaining task, if counter-examples have been found, is to ascertain

whether they arise in the original. This involves defining the path predicate
that leads to each counter-example, expressing it in terms of the original
program variables (that is to say, removing the predicate abstraction, giving,
in the example, a > b and b > a + 1) and determining if any combination of
values for the program variables can satisfy the predicate: if such a
combination, or variable assignment, exists, then the counter-example is a real
one; if not, as in the case given, it is spurious.
This problem of predicate satisfiability is computationally hard; finding
efficient algorithms is one of the central areas of research in model checking.
The focus on counter-examples gives model checking a practical
advantage over traditional proof techniques. Unless a software element was
built with verification in mind (through a constructive method as defined
above), the first attempt to verify it will often fail. With proofs, this failure
doesnt tell us the source of the problem and could actually signal a limitation
of the proof procedure rather than an error in the program. With model checking,
you get a counter-example which directly shows whats wrong.
Model checking has captured considerable attention in recent years, first
in hardware design and then in reactive and real-time systems, for which the
assertions of interest are often expressed in temporal logic.
29
30
DEPENDABLE SOFTWARE
8 DESIGN BY CONTRACT
The goal of developing software to support full proofs of correctness
properties is, as noted, desirable but still unrealistic for most projects. Even a
short brush with program proving methods suggests, however, that more rigor
can be highly beneficial to software quality. The techniques of Design by
Contract go in this direction and deliver part of the corresponding benefits
without requiring the full formality of proof-directed development.
The discussion of proofs introduced Eiffel notations such as
require assertion
-- A routine precondition
ensure assertion
-- A routine postcondition
associated with individual routines. They are examples of contract elements

which specify abstract semantic properties of program constructs. Contracts
apply in particular to:
Individual routines: precondition, stating the condition under which a

routine is applicable; postcondition, stating what condition it will
guarantee in return when it terminates.
In object-oriented programming, classes: class invariant, stating

consistency conditions that must hold whenever an object is in a stable
state. For example, the invariant for a paragraph class in a text
processing system may state that the total length of letters and spaces is
equal to the paragraph width. Every routine that can modify an instance
of the class may assume the class invariant on entry (in addition to its
precondition) and must restore it on exit (in addition to ensuring
its postcondition).
Loops: invariant and (integer) variant as discussed above.
Individual instructions: assert or check constructs.
The discipline of Design by Contract [53] [57] [67] gives a central role to these
mechanisms in software development. It views the overall process of building
a system as defining a multitude of relationships between client and
supplier modules, each specified through a contract in the same manner as
relationships between companies in the commercial world.
The benefits of such a method, if carried systematically, extend
throughout the lifecycle, supporting the goal of seamlessness discussed earlier:
Contracts can be used to express requirements and specifications in a

precise yet understandable way, preferable to pure bubbles and arrows
notations, although of course they can be displayed graphically too.
8 DESIGN BY CONTRACT
The method is also a powerful guide to design and implementation,

helping developers to understand better the precise reason and context for
every module they produce, and as a consequence to get the module right.
Contracts serve as a documentation mechanism: the contract view of

a class, which discards implementation-dependent elements but retains
externally relevant elements and in particular preconditions,
postconditions and class invariants, often provides just the right form of
documentation for software elements, especially reusable components:
precise enough thanks to the contracts; abstract enough thanks to the
removal of implementation properties; extracted from the program text,
and hence having a better chance of being up to date (at least one major
software disaster was traced [41] to a software element whose
specification had changed, unbeknownst to the developers who reused it);
cheap to produce, since this form of documentation can be generated by
tools from the source text, rather than written separately; and multipurpose, since the output can be tuned to any appropriate format such as
HTML. Eiffel environments such as EiffelStudio produce such views
[30], which serve as the basic form of software documentation.
Contracts are also useful for managers to understand the software at a

high level of abstraction, and as a tool to control maintenance.
In object-oriented programming, contracts provide a framework for the

proper use of inheritance, by allowing developers to specify the semantic
framework within which routines may be further refined in descendant
classes. This is connected with the preceding comment about
management, since a consequence is to allow a manager to check that
refinements to an design are consistent with its original intent, which may
have been defined by the top designers in the organization and expressed
in the form of contracts.
Most visibly, contracts are a testing and debugging mechanism. Since an

execution that violates an assertion always signals a bug, turning on
contract monitoring during development provides a remarkable
technique for identifying bugs. This idea is pursued further by some of
the tools cited in the discussion of testing below.
Design by Contract mechanisms are integrated in the design of the Eiffel

language [52] [28] and a key part of the practice of the associated method.
Dozens of contract extensions have been proposed for other programming
languages (as well as UML [80]), including many designs such as JML [43]
for Java and the Spec# extension of C# [5].
31
32
DEPENDABLE SOFTWARE
9 TESTING
Testing [70] [8] is the most widely used form of program verification, and still
for many teams essentially the only one. In academic circles testing has long
suffered from a famous comment [23] that (because of the astronomical
number of possible states) testing can only show the presence of bugs, but
never to show their absence. In retrospect its hard to find a rational
explanation for why this comment ever detracted anyone from the importance
of tests, since it in no way disproves the usefulness of testing: finding bugs is
a very important task of software development. All it indicates is that we
should understand that finding bugs is indeed the sole purpose of testing, and
not delude ourselves that test results directly reflect the level of quality of a
product under development.
Components of a test
Successful testing relies on a test plan: a strategy, expressed in a document,
describing choices for the tasks of the testing process. These tasks include:
Determining which parts to test.
Finding the appropriate input values to exercise.
Determining the expected properties of the results (known as oracles).

Input values and the associated oracles together make up test cases, the
collection of which constitutes a test suite.
Instrumenting the software to run the tests (rather than perform its normal
operation, or in addition to it); this is known as building a test harness,
which may involve test drivers to solicit specific parts to be tested, and
stubs to stand for parts of the system that will not be tested but need a
placeholder when other parts call them.
Running the software on the selected inputs.
Comparing the outputs and behavior to the oracles.
Recording the test data (test cases, oracles, outputs) for future re-testing
of the system, in particular regression testing, the task of verifying that
previously corrected errors have not reappeared.
In addition there will be a phase of correction of the errors uncovered by the

test, but in line with the above observations this is not part of testing in the
strict sense.
9 TESTING
Kinds of test
One may classify tests with respect to their scope (this was used in the earlier
description of the V model of the lifecycle):
A unit test covers a module of the software.
Integration test covers a complete cluster or subsystem.
A system test covers the complete delivery.
User Acceptance Testing involves the participation of the recipients of the

system (in addition to the developers, responsible for the preceding
variants) to determine whether they are satisfied with the delivery.
Business Confidence Testing is further testing with the users, in

conditions as close as possible to the real operating environment.
An orthogonal classification addresses what is being tested:
Functional testing: whether the system fulfills the functions defined in

the specification.
Performance testing: its use of resources.
Stress testing: its behavior under extreme conditions, such as heavy

user load.
Yet another dimension is intent: testing can be fault-directed to find

deficiencies but also (despite the above warnings), conformance-directed to
estimate satisfaction of desired properties, or acceptance testing for users to
decide whether to approve the product. Regression testing, as noted, re-runs
tests corresponding to previously identified errors; surprisingly to the layman,
errors have a knack for surging back into the software, sometimes repeatedly,
long after they were thought corrected.
The testing technique, in particular the construction of test suites, can be:
Black-box: based on knowledge of the systems specification only.
White-box: based on knowledge of the code, which makes it possible for

example to try to exercise as much of that code as possible.
Observing the state of the art in software testing suggests that four issues are
critical: managing the test process; estimating the quality of test suites;
devising oracles; and the toughest generating test cases automatically.
33
34
DEPENDABLE SOFTWARE
Managing the testing process

Test management has been made easier through the appearance of testing
frameworks such as JUnit [42] and Gobo Eiffel Test [7] which record test
harnesses to allow running the tests automatically. This removes a considerable
part of the burden of testing and is important for regression testing.
An example of a framework for regression testing of a compiler,
incorporating every bug ever found since 1991, is EiffelWeasel [29]. Such
automated testing require a solid multi-process infrastructure, to ensure for
example that if a test run causes a crash the testing process doesnt also crash
but records the problem and moves on to the next test.
Estimating test quality

Being able to estimate the quality of a test suite is essential in particular to
know when to stop testing. The techniques are different for white-box and
black-box testing.
With white-box testing it is possible to define various levels of coverage,
each assuming the preceding ones: instruction coverage, ensuring that through
the execution of the selected test cases every instruction is executed at least
once; branch coverage, where every boolean condition tests at least once to true
and once to false; condition coverage, where this is also the case for boolean
sub-expressions; path coverage, for which every path has been taken; loop
coverage, where each loop body has been executed at least n times for set n.
Another technique for measuring test suite quality in white-box
approaches is mutation testing [79]. Starting with a program that passes its
test suite, this consists of making modifications similar, if possible, to the
kind of errors that programmers would make to the program, and running
the tests again. If a mutant program still passes the tests, this indicates (once
you have made sure the mutant is not equivalent to the original, in other words,
the changes are meaningful) that the tests were not sufficient. Mutation testing
is an active area of research [71]; one of the challenges is to use appropriate
mutation operators, to ensure diversity of the mutants.
With black-box testing the previous techniques are not available since
they assume access to the source code to set up the test plan. It is possible to
define notions of specification coverage to estimate whether the tests have
exercised the various cases listed in the specification; if contracts are present,
this will mean analyzing the various cases listed in the preconditions. Partition
testing [81] is the general name for techniques (black- or white-box) that split
the input domain into representative subsets, with the implication that any test
suite must cover all the subsets.
9 TESTING
Defining oracles
An oracle, allowing interpretation of testing results, provides a decision criterion
for accepting or rejecting the result of a test. The preparation of oracles can be as
much work as the rest of the test plan. The best solution that can be recommended
is to rely on contracts: any functional property of a software system (with the
possible exception of some user-interface properties for which human assessment
may be required) can be expressed as a routine postcondition or a class invariant.
These assertions can be included in the test harness, but it is of course
best, as noted in the discussion of Design by Contract, to make them an
integral part of the software to be tested as it is developed; they will then
provide the other benefits cited, such as aid to design and built-in
documentation, and will facilitate regression testing.
Test case generation

The last of the four critical issues listed, test case generation, is probably the
toughest; automatic generation in particular. Even though we cant ever get
close to exhaustive testing, we want the test process to cover as many cases as
possible, and especially to make sure they are representative of the various
potential program executions as can be assessed in white-box testing by
coverage measures and mutation, but needs to be sought in any form of testing.
For any realistic program, manually prepared tests will never cover
enough cases; in addition, they are tedious to prepare. Hence the work on
automatic test case generation, which tries to produce as many representative
test cases as possible, typically working from specifications only (black-box).
Two tools in this area are Korat for JML [13] and AutoTest for Eiffel [15]
(which draws on the advantage that contracts being native to Eiffel
existing Eiffel software is typically equipped with large numbers of assertions,
so that AutoTest can be run on software as is, and indeed has already uncovered
a significant number of problems in existing programs and libraries).
Manual tests, which benefit from human insight, remain indispensable. The
two kinds are complementary: manual tests are good at depth, automatically
generated tests at breadth. In particular, any run that ever uncovered a bug,
whether through manual or automatic techniques, should become part of the
regression test suite. AutoTest integrates manual tests and regression tests within
the automatic test case generation and execution framework [44].
Automatic test case generation needs a strategy for selecting inputs.
Contrary to intuition, random testing [34], which selects test data randomly
from the input domain, can be an effective strategy if tuned to ensure a
reasonably even distribution over that domain, a policy known as adaptive
random testing [14] which has so far been applied to integers and other simple
values (for which a clear notion of distance exists, so that even distribution
is immediately meaningful). Recent work [16] extends the idea to objectoriented programming by defining a notion of object distance.
35
36
DEPENDABLE SOFTWARE
10 CONCLUSION
This survey has taken a broad sweep across many techniques that all have
something to contribute to the aim of software reliability. While it has stayed
away from the gloomy picture of the state of the industry which seems to be
de rigueur in discussions of this topic, and is not justified given the
considerable amount of quality-enhancing ideas, techniques and tools that are
available today and the considerable amount of good work currently in
progress, it cannot fail to note as a conclusion that the industry could do much
more to take advantage of all these efforts and results.
There is not enough of a reliability culture in the software world; too
often, the order of concerns is cost, then deadlines, then quality. It is time to
reassess priorities.
Acknowledgments
The material in this chapter derives in part from the slides for an ETH industry
course on Testing and Software Quality Assurance prepared with the help of
Ilinca Ciupa, Andreas Leitner and Bernd Schoeller. The discussion of CMMI
benefited from the work of Peter Kolb in the preparation of another ETH
course, Software Engineering for Outsourced and Offshored Development.
Bernd Schoeller and Ilinca Ciupa provided important comments on the draft.
Design by Contract is a trademark of Eiffel Software.
The context for this survey was provided by the Hasler Foundations
grant for our SCOOP work in the DICS project. We are very grateful for the
opportunities that the grant and the project have provided, in particular for the
experience gained in the two DICS workshops in 2004 and 2005.
REFERENCES
Note: All URLs listed were active in April 2006.
[1] Algirdas Avizienis, Jean-Claude Laprie and Brian Randell: Fundamental
Concepts of Dependability, in Proceedings of Third Information Survivability
Report, October 2000, pages 7-12, available among other places at
citeseer.ist.psu.edu/article/avizienis01fundamental.html.
[2] Ralph Back: A Calculus of Refinements for Program Derivations, in Acta
Informatica, vol. 25, 1988, pages 593-624, available at crest.cs.abo.fi/publications/
public/1988/ACalculusOfRefinementsForProgramDerivationsA.pdf.
37
10 REFERENCES
[3] Thomas Ball and Sriram K. Rajamani: Automatically Validating Temporal

Safety Properties of Interfaces, in SPIN 2001, Proceedings of Workshop on
Model Checking of Software, Lecture Notes in Computer Science 2057,
Springer-Verlag, May 2001, pages 103-122, available at tinyurl.com/qrm9m.
[4] Mike Barnett, Robert DeLine, Manuel Fhndrich, K. Rustan M. Leino,
Wolfram Schulte: Verification of object-oriented programs with invariants, in
Journal of Object Technology, vol. 3, no. 6, Special issue: ECOOP 2003
workshop on Formal Techniques for Java-like Programs, June 2004, pages 2756, available at www.jot.fm/issues/issue_2004_06/article2.
[5] Mike Barnett, K. Rustan M. Leino and Wolfram Schulte: The Spec#
Programming System: An Overview, in CASSIS 2004: Construction and
Analysis of Safe, Secure Interoperable Smart devices, Lecture Notes in
Computer Science 3362, Springer-Verlag, 2004, available at
research.microsoft.com/specsharp/papers/krml136.pdf; see also other Spec#
papers at research.microsoft.com/specsharp/.
[6] Kent Beck and Cynthia Andres: Extreme Programming Explained:
Embrace Change. 2nd edition, Addison-Wesley, 2004.
[7] ric Bezault: Gobo Eiffel Test, online
www.gobosoft.com/eiffel/gobo/getest/index.html.
documentation
at
[8] Robert Binder: Testing Object-Oriented Systems: Models, Patterns, and

Tools, Addison-Wesley, 1999.
[9] Bruno Blanchet, Patrick Cousot, Radhia Cousot, Jrme Feret, Laurent
Mauborgne, Antoine Min, David Monniaux and Xavier Rival: ASTRE: A
Static Analyzer for Large Safety-Critical Software, in Applied Deductive
Verification, Dagstuhl Seminar 3451, November 2003, available at
www.di.ens.fr/~cousot/COUSOTtalks/Dagstuhl-3451-2003.shtml. See also
ASTRE page at www.astree.ens.fr.
[10] Barry W. Boehm: Software Engineering Economics, Prentice Hall, 1981.
[11] Barry W. Boehm: A Spiral Model of Software Development and
Enhancement, in Computer (IEEE), vol. 21, no. 5, May 1988, pages 61-72.
[12] Barry W. Boehm et al.: Software Cost Estimation with COCOMO II,
Prentice Hall, 2000.
[13] Chandrasekhar Boyapati, Sarfraz Khurshid and Darko Marinov: Korat:
Automated Testing Based on Java Predicates, in Proceedings of the 2002
International Symposium on Software Testing and Analysis (ISSTA), Rome,
July 22--24, 2002, available at tinyurl.com/qwwd3.
38
DEPENDABLE SOFTWARE
[14] T.Y. Chen, H. Leung and I.K. Mak: Adaptive random testing, in Advances
in Science - ASIAN 2004: Higher-Level Decision Making, 9th Asian
Computing Science Conference, ed. Michael J. Maher, Lecture Notes in
Computer Science 3321, Springer-Verlag, 2004, available at
tinyurl.com/lpxn5.
[15] Ilinca Ciupa and Andreas Leitner: Automated Testing Based on Design by
Contract, in Proceedings of Net.ObjectsDays 2005, 6th Annual Conference on
Object-Oriented and Internet-Based Technologies, Concepts and
Applications for a Networked World, 2005, pages 545-557, available at
se.ethz.ch/people/ciupa/papers/soqua05.pdf. See also AutoTest page at
se.ethz.ch/research/autotest.
[16] Ilinca Ciupa, Andreas Leitner, Manuel Oriol and Bertrand Meyer: Object
Distance and its Application to Adaptive Random testing of Object-Oriented
Programs, submitted for publication, 2006, available at se.ethz.ch/~meyer/
publications/testing/object_distance.pdf.
[17] Edmund M. Clarke Jr., Orna Grumberg and Doron A. Peled: Model
Checking, MIT Press, 1999.
[18] Patrick Cousot: Verification by Abstract Interpretation, in International
Symposium on Verification Theory & Practice Honoring Zohar Mannas 64th
Birthday, ed. Nachum Dershowitz, Lecture Notes in Computer Science 2772,
Springer-Verlag, 2003, pages 243-268.
[19] Michael Cusumano and Richard Selby: Microsoft Secrets, The Free
Press, 1995.
[20] Ole-Johan Dahl, Edsger W. Dijkstra and C.A.R. Hoare: Structured
Programming, Academic Press, 1971.
[21] David L. Detlefs, K. Rustan M. Leino, Greg Nelson, and James B. Saxe:
Extended Static Checking, Research Report 159, Compaq Systems Research
Center, December 1998, available at ftp://gatekeeper.research.compaq.com/
pub/DEC/SRC/research-reports/SRC-159.pdf.
[22] Edsger W. Dijkstra: Go To Statement Considered Harmful, in
Communications of the ACM, Vol. 11, No. 3, March 1968, pages 147-148,
available at www.acm.org/classics/oct95/.
[23] Edsger W. Dijkstra: Notes on Structured Programming, in [20]; original
typescript available at www.cs.utexas.edu/users/EWD/ewd02xx/EWD249.PDF.
[24] Edsger W. Dijkstra: A Discipline of Programming, Prentice Hall, 1978.
[25] Brian J. Dreger: Function Point Analysis, Prentice Hall, 1989.
10 REFERENCES
[26] Paul Dubois, Mark Howard, Bertrand Meyer, Michael Schweitzer and
Emmanuel Stapf: From Calls to Agents, in Journal of Object-Oriented
Programming (JOOP), vol. 12, no. 6, September 1999, available at
se.ethz.ch/~meyer/publications/joop/agent.pdf.
[27] Eclipse pages at www.eclipse.org.
[28] ECMA/ISO: Eiffel: Analysis, Design and Programming Language,
standard ECMA 367, accepted in April 2006 as ISO standard, available at
www.ecma-international.org/publications/standards/Ecma-367.htm.
[29] Eiffel open-source development site at eiffelsoftware.origo.ethz.ch/
index.php/Main_Page.
[30] Eiffel Software: EiffelStudio documentation, online at eiffel.com.
[31] Michael D. Ernst, J. Cockrell, William G. Griswold and David Notkin:
Dynamically Discovering Likely Program Invariants to Support Program
Evolution, in IEEE Transactions on Software Engineering, vol. 27, no. 2,
February 2001, pages 1-25, available at pag.csail.mit.edu/~mernst/pubs/
invariants-tse2001.pdf.
[32] Erich Gamma, Richard Helms, Ralph Johnson and John Vlissides:
Design Patterns, Addison-Wesley, 1994.
[33] Carlo Ghezzi, Mehdi Jazayeri, Dino Mandrioli, Software Engineering,
2nd edition, Prentice Hall, 2003.
[34] Richard Hamlet: Random Testing, in Encyclopedia of Software
Engineering, ed. J. J. Marciniak, 1994, available at tinyurl.com/rcjxg.
[35] Brian Henderson-Sellers: Object-Oriented Metrics: Measures of
Complexity, Prentice Hall, 1995.
[36] Thomas A. Henzinger, Xavier Nicollin, Joseph Sifakis and Sergio
Yovine: Symbolic Model Checking for Real-Time Systems, in Logic in
Computer Science, Proceedings of 7th Symposium in Logics for Computer
Science, Santa Cruz, California, 1992, pages 394-406, available at
tinyurl.com/lb5fm.
[37] C.A.R. Hoare: An axiomatic basis for computer programming, in
Communications of the ACM, Vol. 12, no. 10, October 1969, pages 576 - 580,
available at tinyurl.com/ory2s.
[38] C.A.R. Hoare and Jayadev Misra: Verified Software: Theories, Tools,
Experiments, Vision of a Grand Challenge Project, October 2005, foundation
paper for the VSTTE conference [77], available at vstte.ethz.ch/pdfs/vsttehoare-misra.pdf.
[39] IFIP Working Group 10.4 on dependable computing and fault tolerance:
home page at www.dependability.org.
39
40
DEPENDABLE SOFTWARE
[40] Michael Jackson: Problem Frames: Analysing and Structuring Software

Development Problems, Addison-Wesley, 2001.
[41] Jean-Marc Jzquel and Bertrand Meyer: Design by Contract: The
Lessons of Ariane, in Computer (IEEE), vol. 30, no. 1, January 1997, pages
129-130, available at archive.eiffel.com/doc/manuals/technology/contract/
ariane/page.html.
[42] JUnit pages at SourceForge: junit.sourceforge.net.
[43] Gary T. Leavens and Yoonsik Cheon: Design by Contract with JML
(Draft), at ftp://ftp.cs.iastate.edu/pub/leavens/JML/jmldbc.pdf; see also other
JML papers at www.cs.iastate.edu/~leavens/JML/.
[44] Andreas Leitner, Ilinca Ciupa, Bertrand Meyer and Mark Howard:
Reconciling Manual and Automated Testing: The AutoTest Experience,
submitted for publication, 2006.
[45] Nancy G. Leveson: System Safety in Computer-Controlled Automotive
Systems, SAE Congress, March 2000, available at sunnyday.mit.edu/papers/
sae.pdf.
[46] Michael R. Lyu (ed.): Handbook of Software Reliability Engineering,
IEEE Computer Society Press and McGraw-Hill, 1995; also available online
at www.cse.cuhk.edu.hk/~lyu/book/reliability/.
[47] Zohar Manna and Amir Pnueli: The temporal logic of reactive and
concurrent systems, Springer-Verlag, 1992.
[48] Thomas J. McCabe: A Complexity Measure, in IEEE Transactions on
Software Engineering, vol. 2, no. 4, December 1976, pages 308-320.
[49] Thomas J. McCabe and Charles W. Butler: Design Complexity
Measurement and Testing, in Communications of the ACM, vol. 32, no. 12,
December 1989, pages 1415-1425.
[50] Bertrand Meyer: Introduction to the Theory of Programming Languages,
[51] Bertrand Meyer, The New Culture of Software Development: Reflections
on the Practice of Object-Oriented Design, in Advances in Object-Oriented
Software Engineering, eds. D. Mandrioli, B. Meyer, Prentice Hall, 1991.
[52] Bertrand Meyer: Eiffel: The Language, 2nd printing, Prentice Hall, 1992.
[53] Bertrand Meyer: Applying Design by Contract", in Computer (IEEE),
25, 10, October 1992, pages 40-51.
[54] Bertrand Meyer: Object Success, Prentice Hall, 1995.
[55] Bertrand Meyer: Practice to Perfect: The Quality First Model, in
Computer (IEEE), May 1997, pages 102-106, available at se.ethz.ch/~meyer/
publications/computer/quality_first.pdf.
10 REFERENCES
[56] Bertrand Meyer: UML: The Positive Spin, in American Programmer, 1997,
available at archive.eiffel.com/doc/manuals/technology/bmarticles/uml/page.html.
[57] Bertrand Meyer: Object-Oriented Software Construction, 2nd edition,
[58] Bertrand Meyer, Christine Mingins and Heinz Schmidt: Providing
Trusted Components to the Industry, in Computer (IEEE), vol. 31, no. 5, May
1998, pages 104-105, available at se.ethz.ch/~meyer/publications/computer/
trusted.pdf.
[59] Bertrand Meyer: The Role of Object-Oriented Metrics, in Computer
(IEEE), vol. 31, no. 11, November 1998, pages 123-125, available at
se.ethz.ch/~meyer/publications/computer/metrics.
[60] Bertrand Meyer, Every Little Bit Counts: Towards Reliable Software, in
Computer (IEEE_, vol. 32, no. 11, November 1999, pages 131-133, available
at se.ethz.ch/~meyer/publications/computer/reliable.pdf.
[61] Bertrand Meyer: The Grand Challenge of Trusted Components, in ICSE
25 (International Conference on Software Engineering, Portland, Oregon,
May 2003), IEEE Computer Press, 2003.
[62] Bertrand Meyer: The Power of Abstraction, Reuse and Simplicity: An
Object-Oriented Library for Event-Driven Design, in From ObjectOrientation to Formal Methods: Essays in Memory of Ole-Johan Dahl, eds.
Olaf Owe, Stein Krogdahl, Tom Lyche, Lecture Notes in Computer Science
2635, Springer-Verlag, 2004, pages 236-271, available at se.ethz.ch/~meyer/
publications/lncs/events.pdf.
[63] Bertrand Meyer: Offshore Development: The Unspoken Revolution in
Software Engineering, in Computer (IEEE), January 2006, pages 122-124,
available at se.ethz.ch/~meyer/publications/computer/outsourcing.pdf.
[64] Bertrand Meyer: What will remain of Extreme Programming?,
in EiffelWorld, Vol. 5, no. 2, February 2006, available at www.eiffel.com/
general/monthly_column/2006/February.html.
[65] Bertrand Meyer and Karine Arnout: Componentization: the Visitor
Example, to appear in Computer (IEEE), 2006, draft available at
se.ethz.ch/~meyer/publications/computer/visitor.pdf.
[66] Microsoft: Visual Studio pages at msdn.microsoft.com/vstudio.
[67] Richard Mitchell and Jim McKim: Design by Contract by Example,
Addison-Wesley, 2001.
[68] Carroll Morgan: Programming from Specifications, 2nd edition, Prentice
Hall, 1994, available at web.comlab.ox.ac.uk/oucl/publications/books/PfS/.
[69] John Musa: Software Reliability Engineering, 2nd edition, McGraw-Hill, 1998.
41
42
DEPENDABLE SOFTWARE
[70] Glenford J. Myers, Corey Sandler, Tom Badgett and Todd M. Thomas:
The Art of Software Testing, 2nd edition, Wiley, 2004.
[71] Jeff Offutt: Mutation testing papers at www.ise.gmu.edu/~ofut/rsrch/
mut.html.
[72] John Pincus: presentations (mostly PowerPoint slides) on PREfix and
PREfast at research.microsoft.com/users/jpincus/.
[73] Eric Raymond: The Cathedral and the Bazaar: Musings on Linux and
Open Source by an Accidental Revolutionary, O Reilly, 1999; earlier version
available at www.firstmonday.org/issues/issue3_3/raymond/.
[74] Software Engineering Institute, CMMI site, available at
www.sei.cmu.edu/cmmi.
[75] Matt Stephens and Doug Rosenberg: Extreme Programming Refactored:
The Case Against XP, aPress, 2003.
[76] Axel van Lamsweerde: Goal-Oriented Requirements Engineering: A
Guided Tour, in Proceedings of the 5th IEEE International Symposium on
Requirements Engineering, August 2001, available at tinyurl.com/mscpj.
[77] Verified Software: Theories, Tools, Experiments: International IFIP
conference, ETH Zurich, October 2005, see VSTTE conference site at
vstte.ethz.ch.
[78] John Viega: The Myth of Open-Source Security, 2000, available at
www.developer.com/tech/article.php/626641; follow-up article, Open-Source
Security:
Still
at
Myth,
September
2004,
available
at
www.onlamp.com/pub/a/security/2004/09/16/open_source_security_myths.html.
[79] Jeffrey M. Voas and Gary McGraw: Software Fault Injection: Inoculating
Programs Against Errors, Wiley, 1998.
[80] Jos Warmer and Anneke Kleppe: The Object Constraint Language:
Getting Your Models Ready for MDA, 2nd edition, Addison-Wesley, 2003.
[81] Elaine J. Weyuker and Bingchiang Jeng: Analyzing Partition Testing
Strategies, in IEEE Transactions on Software Engineering, vol. 17, no. 9, July
1991, pp. 97-108.
[82] Wikipedia: entry Mars Climate Orbiter, available at
en.wikipedia.org/wiki/Mars_Climate_Orbiter.
[83] Edward Yourdon: When Good Enough Software Is Best, in Software
(IEEE), vol. 12, no. 3, May 1995, pages 79-81.

Dependability

Uploaded by

Copyright:

Available Formats

Dependability

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dependability

Uploaded by

Copyright:

Available Formats

Dependable software

Bertrand Meyer, ETH Zurich

Cite as follows: Bertrand Meyer, Dependable

2 SCOPE AND TERMINOLOGY

Reliability and dependability

There does exist a definition of dependability [1] from the eponymous

2 SCOPE AND TERMINOLOGY

Correctness, robustness, security

Correctness is a systems ability to perform according to its specification

Robustness is a systems ability to prevent damage in cases of erroneous

Security is a systems ability to prevent damage in cases of hostile use

Product and process

The products of software

A fault is a departure of the software product from the properties it should

An error is a wrong human decision made during the construction of the

2 SCOPE AND TERMINOLOGY

Verification and validation

Verification is internal assessment of the consistency of the product,

Validation is relative assessment of a product vis--vis another that

Criteria for classifying approaches to software reliability

A posteriori (assess and correct)

Product- and phase-neutral

Partial (some progress)

The first distinction is cultural almost as much as it is technical. With a priori

An idea can be applicable regardless of technology choices; for example

The Spiral model [11] which focuses on reducing risk in project

The Rational Unified Process, distinguishing four phases, inception,

The Cluster model [51] [57], emphasizing a different form of

The figure shows pictorial representations of some of these models.

Spiral (from [11])

Lifecycle models, illustrated

Whatever their effect on how people actually develop software, the

A distinction between requirements, the recording of user requirements,

Recognition of the importance of Verification and Validation tasks.

Recognition of post-delivery activities such as maintenance, although

The growing emphasis on incrementality in the development process,

Defined: proactive process defined for the organization.

Quantitatively managed: the control mechanisms do not limit themselves

Optimizing: the mechanisms for controlling processes are sufficiently

Short release cycles to get frequent feedback.

Pair programming (two people at a keyboard and terminal).

A general distrust of specification and design: testing is the preferred

Emphasis on programmers welfare.

Small meeting: at most 8 people or so, including the developer of the

The elements under review and any supporting documents must be

The meeting must have a moderator to guide discussions and a secretary

The moderator should not be the developers manager. The intent is to

Another practical impediment to continued quality improvement comes

5 TOOLS AND ENVIRONMENTS

Register a new version of any project element.

Retrieve any previously registered version of any project element.

Register dependencies, both between project elements and between

Construct composite products from their constituents for example,

A significant number of software disasters on record followed from

5 TOOLS AND ENVIRONMENTS

Metrics and models

Integrated development environments

Techniques for information hiding, which protect modules against

Inheritance, allowing the classification and systematic organization of

Genericity, allowing the construction of type-parameterized modules.