University of Southern California Dissertations and Theses

Incremental development productivity decline

(USC Thesis Other)

Incremental development productivity decline

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Incremental Development Productivity Decline

by

Ramin Moazeni

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
Doctor of Philosophy
(Computer Science)

August 2014

Copyright 2014 Ramin Moazeni

ii
DEDICATION
To my parents, Nahid and Mahmoud
To my grandfather, Ali who I lost this year
And to my wife, Laleh

iii
ACKNOWLEDGMENTS
I would like to greatly thank everyone who was a part of my journey. First, I
would like to thank my advisor, Dr. Barry Boehm for guiding and advising me through
the research journey. He was the main inspiration for me to pursue the Ph.D. degree. His
vision and great advices on selecting the right topic and taking the correct approach
helped me through all these years.
I am grateful to my dissertation committee members, Dr. Aiichiro Nakano, and
Dr. Stan Settles, for the additional guidance and suggestions in completing my
dissertation. Also, I would like to extend my thanks to the professors that were part of the
Ph.D. qualifying exam committee, Dr. William GJ Halfond and Dr. Richard Selby, for
comments and feedback on my topic proposal.
I would like to thank all my friends at USC Computer Science department and
members of the Center for Systems and Software Engineering Lab for their valuable
friendship which made my studies joyful. I would also like to specially thank my
colleague Daniel Link that our working together has not only inspired me, but has helped
me finish my dissertation in a fun way. He is a true friend I can really rely and build on.
I wish to express gratitude to my family, my parents Nahid and Mahmoud, for
their great deal of support throughout my PhD studies. They’ve taught me about hard
work and self-respect, about persistence and about how to be independent. They both
were great role models of resilience, strength and character.
Last but not least, I adore the love and support from my dear and loving wife,
Laleh, for sincere encouragement and inspiration throughout my research work and
lifting me uphill this phase of life. She was always there for me through all the difficult
iv
times. Without her it would have been unquestionably much harder to finish a PhD while
having a fruitful and challenging career.
v
TABLE OF CONTENTS
1 Introduction 1
1.1 Background 1
1.2 Central Hypotheses 3
1.3 Iterative and Incremental Development 4
1.3.1 Incremental Development Definition 4
1.3.2 Incremental Development Types 5
1.4 Definitions 8
1.4.1 Increments and builds 8
1.4.2 Productivity 8
1.4.3 IDPD 10
1.4.4 IDPD factor 11
1.4.5 Categories of projects (IDPD types) 12
2 Background and Related Work 19
2.1 Major Cost Estimation Models 19
2.1.1 Cost Estimation Considerations 19
2.1.2 COCOMO II Cost Drivers and Scale Factors 22
2.2 Cost Model for Maintenance and Reuse 24
2.3 Laws of Software Evolution 26
2.3.1 Lehman’s System Types 26
2.3.2 Lehman’s Laws of Software Evolution 27
2.3.3 Discussion of the individual Laws of Software Evolution 29
3 Research Approach 39
3.1 Behavioral Analysis 39
vi
3.2 Data Collection and Analysis 39
3.2.1 Controlled Experiment 42
3.2.2 Open Source Projects 43
3.2.3 Data Collection Challenges 44
3.3 Contextual Analysis 45
3.4 Statistical Analysis 46
3.4.1 Linear Correlation 46
3.4.2 Tailed Pair T-Test 46
3.4.3 One-Way ANOVA 48
4 Analysis and Representation of Data 49
4.1 Summary of Data Collection and Selection 49
4.2 Productivity Decline: Hypothesis 1 51
4.2.1 Case Study 1: Quality Management Platform 52
4.2.2 Case Study 2: XP1 and XP2 54
4.2.3 Statistical Significance 56
4.3 Build-to-Build Behavior: Hypothesis 2 61
4.4 Domain Variation: Hypothesis 3 63
4.5 COCOMO II and IDPD 65
4.5.1 Cost Estimation 65
4.6 IDPD and COCOMO II 67
4.6.1 COCOMO II cost drivers and scale factors 67
4.6.2 How IDPD influences parameters 71
4.6.3 Potential New Cost Estimation Parameters 71
Time between increments 72
Defect number and complexity in increments 73
vii
Comparative size of current increment 73
Coherence of the project 74
Evaluated new Cost Drivers 74
4.7 Threats to Validity 75
4.7.1 Internal validity considerations 75
4.7.2 External validity considerations 76
5 Incremental Development Cost Estimation 77
5.1 Incremental COCOMO II 78
5.1.1 Incremental COCOMO Estimation Steps 78
5.1.2 Incremental Development Estimation Considerations 80
5.2 Constructive Incremental Cost Model 80
5.2.1 Software Engineering Reasons for Multi-Build, Overlapping RUP Cycles 80
5.3 Incremental Development Cost Estimation using COINCOMO 82
5.3.1 COINCOMO Cost Estimation Steps 85
6 The COINCOMO Tool 87
6.1 COINCOMO Terminology 88
6.2 The COINCOMO Architecture 89
7 Conclusion and Future Work 93
7.1 General Conclusions 93
7.2 Summary of Contributions 94
7.3 Future Work 95
Appendix A: The COINCOMO User Interface 100
Appendix B: Working with COINCOMO 113

viii
LIST OF TABLES
Table 1. IDPD Effort Drivers (Madachy R. , Boehm, Clark, Tan, & Rosa, 2011) ........... 11
Table 2: IDPD intensity by category ................................................................................ 18
Table 3: Effect of COCOMO II effort multipliers on IDPD ............................................ 23
Table 4: Effect of COCOMO II Scale Factors on IDPD .................................................. 24
Table 5: Core collected data attributes.............................................................................. 41
Table 6: Summary of data sources .................................................................................... 44
Table 7: Incremental project domains ............................................................................... 45
Table 8: COCOMO II Driver ratings ................................................................................ 53
Table 9. Correlation coefficient (with p-value) between build-to-build productivity and
number increments .................................................................................................... 57
Table 10. Confidence Value, t-stat value, and p-value of 2 tailed pair t-test .................... 62
Table 11. One-way ANOVA results for Hypothesis 3 ..................................................... 64
Table 12. COCOMO II Cost Drivers and Scale Factors expected to have a decrease in
effort .......................................................................................................................... 66
Table 13. COCOMO II Cost Drivers and Scale Factors expected to have no specific trend
in effort...................................................................................................................... 68
Table 14. P-value of weighted multiple regressions between cost drivers and productivity
................................................................................................................................... 69
Table 15. P-value (<0.08) of regression between cost drivers and productivity for 4 DR
(directed research) projects ....................................................................................... 70
Table 16. Weighted regression between cost drivers and productivity (only for cost
drivers with correlation> 0.7) ................................................................................... 71
ix
Table 17. Correlation of Code Size Metrics to Increments .............................................. 74
Table 18. Comparison of cost estimation models ............................................................. 77

x
LIST OF FIGURES
Figure 1. Incremental development types ........................................................................... 7
Figure 2. Effect of IDPD on number of builds ................................................................. 13
Figure 3. Productivity for all projects on a logarithmic scale ........................................... 50
Figure 4. Average SOLC of projects ................................................................................ 50
Figure 5. Histograms showing number of increments across projects ............................. 51
Figure 6: QMP productivity across increments ................................................................ 54
Figure 7. XP1 productivity across increments .................................................................. 55
Figure 8. XP 2 Productivity across Increments ................................................................ 56
Figure 9. Normalized productivity by increments for 8 projects ...................................... 58
Figure 10. Normalized productivity trend lines ................................................................ 58
Figure 11. Overview of productivity trends across projects ............................................. 60
Figure 12. P-value of 2 tailed t-test across projects .......................................................... 63
Figure 13: Averages of Average IDPD ............................................................................. 64
Figure 14: IDPD Distribution over Domains .................................................................... 64
Figure 15: Average IDPD over Domains.......................................................................... 64
Figure 16. Overlap across Builds ...................................................................................... 82
Figure 17. RUP with MBASE Anchor Point .................................................................... 83
Figure 18. Overlap across builds with carry forward ....................................................... 84
Figure 19: COINCOMO Element Relationship diagram .................................................. 90
Figure 20: COINCOMO Component estimation .............................................................. 91
Figure 21: COPSEMO calculation.................................................................................... 91
Figure 22: COINCOMO support for multibuild estimation ............................................. 92
xi
ABSTRACT
Software production is on the critical path of increasingly many program abilities
to deliver effective operational capabilities. This is due to the number, complexity,
independence, interdependence, and software-intensiveness of their success-critical
components and interfaces. The estimation parameters and knowledge bases of current
software estimation tools are generally good for stable, standalone, single increment
development. However, they do not fully account for the degrees of program and
software dynamism, incrementality, coordination, complexity, and integration. These
phenomena tend to decrease software productivity relative to the cost model estimates
made for the individual software components and for the overall systems, but it is
difficult to estimate by how much.
Incremental software development generally involves either adding, modifying, or
deleting parts of the code in the previous increments. This means that if a useful system is
to be built, the maintenance that will have to go into previous increments will take away
productivity from the later ones.
This research tests hypotheses about a phenomenon called Incremental
Development Productivity Decline (IDPD) that may be more or less present in
incremental software projects of various categories.
Incremental models are now being used by many organizations in order to reduce
development risks while trying to deliver releases of the product on time. It has become
the most common method of software development with characteristics that influence the
productivity of projects.
xii
Different ways of measuring productivity are presented and evaluated in order to
come to a definition or set of definitions that is suitable to these categories of projects.
Data from several sources has been collected and analyzed, and hypotheses tested about
the degree of IDPD and its variation by increment and category. The results indicated the
existence of an IDPD phenomenon, that its magnitude varies by application category, but
that it tended to vary from increment to increment.

1

1 Introduction
1.1 Background
There have been indications on individual projects that economies of scale do not apply to
software increments as they do to increased unit numbers of hardware. While hardware productivity
benefits from an increase in units, the “unit cost of later software increments tend to increase, due to
previous-increment breakage and usage feedback, and due to increased integration and test effort.
Thus, using hardware-driven or traditional software-driven estimation methods for later increments
will lead to underestimates and over-runs in both cost and schedule” (Madachy R. , Boehm, Clark,
Tan, & Rosa, 2011).
The following example of a large defense software system is given in (Madachy R. , Boehm,
Clark, Tan, & Rosa, 2011) (SLOC stands for “Source Lines of Code):
• 5 builds, 7 years, $100M cost
• Build 1 productivity over 300 SLOC/person-month
• Build 5 productivity less than 150 SLOC/person-month (including build 1-4 breakage,
integration, and rework)
• 318% change in requirements across all builds
A constant IDPD factor (i.e. the reduction in productivity between two subsequent builds) of
20% over the course of four increments would result in a reduction of productivity to about 48% of
the original one (Madachy R. , Boehm, Clark, Tan, & Rosa, 2011). At the time that (Madachy R. ,
Boehm, Clark, Tan, & Rosa, 2011) was written, there were indications of similar IDPD phenomena,
but very little data. For example, IDPD factors of 14% (with variations, in a smaller software
2

system) had been found. Large schedule slippages had been found for later increments of large
commercial software applications, such as Microsoft’s Word for Windows (Gill & Iansiti, 1994),
Windows Vista and others (Elssamadisy & Schalliol, 2002) (Madachy R. , Boehm, Clark, Tan, &
Rosa, 2011), but specific data on the quantities were not available.
Intuitively, an IDPD for software increments seems reasonable. Each increment adds to the
body of the existing code for the project. When each increment depends on the previous increments,
the amount of code that increment has to be compatible with and integrated with increases.
Additionally, being dependent on previous increments, the developers will uncover previously
unknown errors in the existing code and have to fix them. They will also have to rework parts of it,
all on the budgets available for the current increment.
On the other hand, though, depending on the characteristics of a given project, the
productivity may rise after several increments with declining productivity. This productivity gain
might happen for several reasons: learning how to deal with the system (e.g. through figuring out
the architecture); from later increments having less complex requirements; or later increments that
may not depend on all parts of earlier ones.
Given the issues and considerations about incremental development, it is instructive to look
at a typical incremental development project and compare productivity-related situations that will
arise if the increments are developed in parallel, sequentially, or with overlapping timeframes.
The first consideration relevant to this is the degree of dependence between increments.
Conceivably, a given increment can depend on any other given increment: not at all; partially; or
completely.
The earliest part at which work on an increment can start depends on how far along the
previous ones need to be. From a technological standpoint if an increment has no dependencies, it
3

can be started on immediately. There will still need to be considerations on the amount of resources
that can be committed to increments that are developed in parallel. The tradeoff is between the
amounts of resources being put into development versus the time the development will take. Where
dependencies exist, increments have to wait until the previous ones reach to the point that allows
them to go forward. The choice then lies in whether the next increment should be started at that
point or later. The longer the time span covered by the development, the higher the cost and risk of
reworking older increments.
The objective of this research has been to gather sufficient data to be able to test hypotheses
about the existence of an IDPD phenomenon, about its magnitude, about its consistency from
increment to increment.
1.2 Central Hypotheses
In order to describe our central hypotheses, we need to understand what productivity is. The
conventional productivity equation is Output per Input. While this is simple to say, different metrics
have been used to measure this in terms for software development. There is a host of different
output metrics (SLOC, Equivalent Source Lines of Code aka ESLOC, function points, and business
value to name a few), but the available data from the observations limits which metrics can be used.
Typically, in software engineering, productivity is measured in SLOC per person-hour. Formal
definition of productivity is presented in 1.4.3.
The principal research question to be addressed by this research is:
How does the productivity behave over the course of an incremental project? Are there patterns to
it? What are the factors that influence it? Does its behavior vary by class of software?
4

To address the above questions, the following null hypotheses have been formulated to test whether
they can be rejected.
Productivity Decline Hypothesis: In incrementally developed software projects that have
coherence and dependency between their increments, productivity doesn’t decline over their course.
Coherence and dependency between increments means that the result of the project is not merely an
accumulation of different pieces of software that might just as well be developed concurrently
without a loss in overall productivity. An example for such a case would be a set of tools that are
developed independently of each other and sold together.
Build-to-Build Behavior Hypothesis: The rate of productivity decline from increment to
increment is relatively constant. Although some projects and “Laws” (Lehman M. , 1980) (Lehman,
Ramil, Wernick, & Perry, 1997) suggest that there is a statistically invariant percentage of
productivity decline across increments, this may not be the case in general.
Domain Hypothesis: For different domains (IDPD types), the average decline in
productivity over the course of a project doesn’t vary significantly.
1.3 Iterative and Incremental Development
1.3.1 Incremental Development Definition
In order to come to a useful definition of incremental development, it helps to look at the
definition of “increment”: for software, a generally accepted definition of an increment is an
improvement evolving additions to, modifications of, and deletions of existing increments’ code
(Nguyen, 2010).
Incremental development therefore has to be a process by which each successive release
adds value to the product, albeit in different ways. While at first it may look promising to make a
5

distinction between iterative development as a process oriented toward a specific end goal that
merely perfects the product through repetition of a sub-process cycles and incremental development
as one that is more open-ended and adds functionality at each step, that distinction is not tenable
when it is considered that the goal-oriented iterative process also leads to the addition of
functionality at every step (Larman, 2004). Additionally, whether or not a project will be ended at a
certain point can be hard to know a priori.
This dissertation focuses on the cases of incremental development for which the increments
are:
• done in more than one step,
• where each new step adds new functionality, and
• where there is coherence and dependency the new step and the previous steps
Additionally, in order to be considered an increment for this IDPD research, an increment
will have to contribute a significant amount of new functionality and must add a significant amount
of size over the previous one, i.e. it must not be less than a tenth of the size of the previous one and
it must not be just a bug fix of the previous one. Otherwise it is considered part of the previous
increment.
1.3.2 Incremental Development Types
Four models of incremental development have been identified (Boehm & Lane, DoD
Systems Engineering and Management Implications for Evolutionary Acquisition of Major Defense
Systems, 2010) (SEBoK). For the purposes of this research, two of them will be combined into one.

6

Pre-specified Sequential model
In this model, development is split up so initial operational capability can be achieved early.
This is then followed by several pre-planned product improvements. The evolution of the project is
limited by the pre-specified plan.
Evolutionary Sequential model
The second strategy is the Evolutionary Sequential model. It rapidly leads to an initial
operational capability, which is upgraded based on operational experience. Agile software
development fits this model. Any defects of a build will be fixed in the next one.
Boehm and Lane also identify a model called the Evolutionary Overlapped model. It covers the
special case of deferring the next increment until critical enablers such as desired new technologies,
anticipated new commercial product capabilities, or needed funding become available or mature
enough to be added. As far as this research is concerned, this is regarded as a mere variant of the
Evolutionary Sequential model in which the time interval between increments is unknown or harder
to plan for.
Evolutionary Concurrent model
In the Evolutionary Concurrent model, several increments are worked on at the same time in
a staggered way: For example, while part of the team develops one increment, another part already
scopes and architects the next one. This mode of operations can be extended to any number of
increments.
Figure 1 shows the workflow of the four identified incremental types (Boehm & Lane, 2010)
(SEBoK):

7

Time phasing terms: Scoping; Architecting; Developing; Producing; Operating (SADPO)
Pre-specified Sequential: SA; DPO1; DPO2; DPO3; …
Evolutionary Sequential: SADPO1; SADPO2; SADPO3; …
Evolutionary Concurrent: SA; D1 ; PO1…
SA2; D2 ; PO2…
SA3; D3; PO3 …
Pre-specified Sequential

Evolutionary Sequential

Evolutionary Concurrent

Figure 1. Incremental development types
8

1.4 Definitions
1.4.1 Increments and builds
To understand our empirical considerations here, it is important to know the definitions of
“increment” and “build” use in this research. For simplicity, the definitions exclude deleted code,
which was usually insignificant in the projects analyzed.
• A build is a released version of the software that is comprised of all the code that has been
written up to the point the build was released.
• An increment is only the code that has been written for a given increment. This includes new
code and adapted/reworked/fixed code from previous increments.
• A build (at least beyond the third one) therefore contains the code from several previous
increments, however it may be dispersed over the build.
1.4.2 Productivity
Productivity is defined as the ratio between “the amount of goods or services produced” and
“the labor or expense that goes into producing them”. Similarly, in software projects, productivity
can be measured as the ratio between “the amount of software” or code produced to “the labor and
expense of producing it” (Boehm B. , 1981). However, this simple approach doesn’t work too well
in real world scenarios. There are so much more factors that affect productivity measures and
prediction. Boehm (Boehm B. , Improving Software Productivity, 1987) defined productivity as
“the ratio of output units produced per unit of input effort”, for example, it could be lines of source
codes versus person-month effort. Scacchi (Hurley, 1995) (Scacchi, 1991) pointed out that
productivity is comparable when the outputs and inputs are of the same kind, i.e. “a software
9

development effort with productivity 2X is twice as productive as another effort whose productivity
is X”. In the IBM federal study case, Walston and Felix (Walston & Felix, 1977) measured
productivity in terms of number of lines of code produced per person‐hour, and in another IBM
study that was done by Albrecht (Albrecht A. , 1979) (Albrecht & Gaffney, 1983) he used “function
points” to measure productivity. A function point is defined as “a composite measure of a number
of program attributes that are multiplied by weighting factors then added together”. In Albrecht’s
approach, productivity wasn’t simply just a ratio, but he introduced various attributes, which could
potentially have effects on productivity measurement, such as file accesses, function calls, etc. In
TRW’s software cost estimation program, Boehm (Boehm B. , 1981) found that “software cost
drivers are effectively the inverse of productivity drivers”. The whole idea is that drivers that yield a
higher cost will conversely yield a lower productivity. For example, when the staff capacity is high
and the product complexity is low, it will lead to high productivity and low cost software
production. Vosburgh and associates did one of the most substantial studies of large-scale software
productivity (Vosburgh, 1984). They defined and divided productivity drivers according to “the
ability of a software project manager to control them”. They identified two types of factors:
product‐related factors that are not usually controllable by a project manager and production process
related factors that are controllable by managers. Within each type, they also identified more
detailed and specific constraints. Overall, software productivity cannot be understood nor measured
simply as a ratio of the amount of produced codes for some unit of time. It requires a more
systematic and statistical analysis of various production measures, as well as their
inter‐relationships.

10

The conventional productivity formula is:

=

In the case of software development, we generally define productivity as:

=

(SLOC can be replaced by KSLOC as needed.)
The problem with measuring productivity in this way is that it regards all SLOC that are
produced as an equal measure of productivity, without differentiating whether a given amount of
SLOC is a desirable product or just a byproduct of the development team’s effort. For example,
SLOC that are produced in order to fix errors, refactor existing code, generated by and IDE or from
duplicating or cutting and pasting code all count as productive based on this formula. Also for
software maintenance , Nguyen (Nguyen, 2010) found that deleting code often required a good deal
of effort. Since any data collection didn’t find much deleted code, I adopt SLOC added in the new
increment as the definition of productivity as SLOC/Effort in this research.
1.4.3 IDPD
Incremental development productivity decline is a phenomenon in which there is an overall
decline in productivity of the increments as measured in section 4. It should be clarified that for
there to be IDPD, an overall decline includes cases where the decline between increments is not
constant and where there can even be a rise in productivity between two given increments. In such
cases, an average productivity decline can be derived from a negative data fitting trendline.
11

1.4.4 IDPD factor
The IDPD factor is the percentage by which productivity decreases between two given
increments. Since it describes a decrease, it is positive if the productivity decreases and negative if
the productivity increases between the two increments.
A build-to-build IDPD factor of about 16% will cause a reduction from the initial
productivity to 50% during the fifth new build. That such declines can happen in actual projects is
shown by a project that forms the basis of one our case studies, which had an IDPD factor of 14%
(Tan, 2009). Large-scale commercial software projects, such as Word for Windows, Windows Vista
and others have had similar experiences (Elssamadisy & Schalliol, 2002) (Gill & Iansiti, 1994). The
table below shows experience-based IDPD effort drivers and their ranges for a four-build system
(based on the table in (Madachy R. , Boehm, Clark, Tan, & Rosa, 2011)):
Table 1. IDPD Effort Drivers (Madachy R. , Boehm, Clark, Tan, & Rosa, 2011)
Personnel experience gain (variation is due to
different turnover rates)
-5% to -20%
Breakage, maintenance of full code base 20-40%
Diseconomies of scale in development, integration 10-25%
Requirements volatility, user requests 10-25%

Note that personnel experience gain decreases IDPD, while the other drivers listed in the
remaining three rows increase it. Basically, work needs to be done on the code in the previous
increments. But needs to be done on the budget for the current increment.
12

The best case would be 20% more effort per increment when the increasing drivers are
minimized and the decreasing one maximized (20% + 10% + 10% - 20%). This would amount to an
productivity decline of 6% per increment for a four-build system.
The worst case would have the decreasing driver minimized and the increasing one
maximized, which would add up to 85% more effort (40% + 25% + 25% - 5%); for a 4-build
system, the productivity decline per increment would be 23%.
The difference between 6% and 23% may not seem large, but over a significant number of
builds, the difference will be striking. Figure 2 shows how a seemingly small difference in the
productivity decline can stretch out a project to twice as many builds and a factor of 2 schedule
overruns, for a four million SLOC system able to produce one million SLOC in its first increment
(Madachy R. , Boehm, Clark, Tan, & Rosa, 2011). Thus, it is important to understand the IDPD
factor and its influence when doing incremental or evolutionary development. Testing of hypotheses
4 indicates that the magnitude of the IDPD factor may vary by type of application (infrastructure
software having higher IDPDs since it tends to be tightly coupled and touches everything;
applications software having lower IDPDs if it is architected to be loosely coupled), or by recency
of the build (older builds may be more stable). Further data collection and analysis will be very
helpful in improving the understanding of the IDPD factor.
1.4.5 Categories of projects (IDPD types)
Before the different categories of projects and their typical IDPD-related relevance are
considered, it needs to be said that these are only the typical cases and that ultimately any software
project, no matter how big or small, can be done with or without increments. Similarly, any
13

software that would normally fit one category can be treated like a member of another. The ultimate
decision lies with the organization conducting the development.

Figure 2. Effect of IDPD on number of builds
Non-Deployable Support Software
Non-deployable code is code that either cannot be deployed due to its nature or which its
author(s) do not intend to deploy. Typically this is code that is developed ad hoc as auxiliary related
to one or more given projects, but not a part of its deliverables. For turnkey software to be
maintained elsewhere, test and other support software should be part of the deliverables. Examples
include configuration scripts, XML-based configuration settings, compiler scripts, compiler tests
(“Hello World”), tests needed for conditional breakpoints, repository commit scripts and small
internal one-off projects such as scripts to run software at trade shows. Additionally, while it may
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
0 2 4 6 8 10
MSLOC
Build
Effects of IDPD on Number of Builds
0% decline
10% decline
15% decline
20% decline
14

be theoretically possible to deploy some subcategories of this code, the deployed code would have
no value to any project stakeholder outside of the author’s organization.
Due to its small size and its ad hoc character, this kind of code is generally not developed in
increments. Even if it would be done, the increments would not be of any meaningful size that
would allow inferences about major projects.
Infrastructure Software
Infrastructure software is software that is designed to not interact with the user directly, but
to provide facilities and services for other software to run. Examples include operating systems (e.g.
Mac OS, Windows, Linux), virtual machines (e.g. the Java Virtual Machine or the CLR in
Microsoft .NET) and operating system extensions through libraries with APIs (such as DirectX in
Windows) as well as common operating environments. This software can also be more specialized
than the examples given here, such as software providing email, conferencing, geolocation, or
weather geolocation services.
Since the need of users to continue using a given software product that relies on specific
infrastructure software tends to often outlast one or more of that infrastructure software’s major
releases, the compatibility needs of infrastructure software as desired by the customers tend to be
high.
While new, incompatible versions of infrastructure software are released from time to time
alongside new operating systems or separately, it is often possible to have several versions of the
same infrastructure software installed and have applications or the user decide which installed
version is used by a given application (this is true for the Java Virtual Machine and DirectX).

15

Application Software
Application software is the software general users interact with directly in most cases and
which therefore represents the computer experience for them. Examples of incrementally developed
COTS products include web browsers (MS Internet Explorer, Google Chrome, Mozilla Firefox,
Apple Safari) and software that is part of office suites (e.g. Word as part of MS Office), but also
custom-built applications that are created for specialized usage.
Of major relevance for the compatibility needs of a given application is whether or not it is designed
to read and write documents that were created by it or other applications. If so, the application will
require the documents to be compatible with the current and historic data formats of itself and those
other applications.
Any application will have requirements regarding infrastructure software because it will at
least require an operating system to run on. Therefore the customers of such software tend to have
two main compatibility needs: Operating system compatibility (will a given version of the
application still work on a new operating system release?) and document compatibility (will a new
version of the application still be able to work with documents that were created with previous
versions?)
The functionality of COTS applications tends to be incremented in different ways over
different phases, with the following pattern being typical:
Phase 1: Toward the first release
Before the first release, functionality is built up over several non-public unreleased
increments. At some point before the release, the makers of the software may allow external beta
testers to try out the software and give feedback. Beta licenses typically contain clauses that
16

document formats are subject to change at any time and that compatibility is not guaranteed. Since
no customer external to the maker of the application has used the application to create value yet or
at least has no reasonable expectation of compatibility between the pre-release versions and the
finished product, the compatibility needs of the application are nil at this point.
Phase 2: After the first release
From this point on, increments will continue, and there are three ways in which they are
presented (or not) to the outside world:
• Internal builds: These are increments that are neither published nor deployed.
• Bug fixes and minor versions: These are increments that are generally delivered to
customers as free updates. They tend to be named “bug fixes”, “updates” or “service packs”.
Compatibility needs are high because the customers regard versions with the same major
release number or name as one and the same.
• Major releases: These are new full releases of the software that are presented as new
versions, which typically require customers to buy a new license or renew their existing one.
Compatibility needs tend to be intermediate in that it has been established over time that
major releases are the typical point where a change in the document format of an application
is to be expected. This is generally accepted as long as previous document formats can be
read or written to (albeit with some loss of document features when compared to the new
version). Infrastructure compatibility tends to be changed as well in major releases. In most
cases, previous releases are supported for a while and the users of the old versions are given
an upgrade path to the new version.
17

Some applications do not only require infrastructure software to be installed, but also offer
access to their own services through a programmable interface (API). If that is the case, it adds
additional compatibility requirements that are at about the same level as the ones that apply to
documents. The primary IDPD driver for application will be the degree of coupling between
existing and new-increment applications. The closer the coupling, the higher the breakage in the
existing application and higher IDPD.
Therefore overall the compatibility needs of applications are high, but with some flexibility
at periodical intervals.
Platform Software
These are hardware drivers for operating systems. Such drivers interact with the hardware
on one side (e.g. over memory-mapped I/O and hardware interrupts) and with the operating system
on the other side using an interface specified by the operating system. They do not store documents
and have no other APIs.
Since they can be arbitrarily rewritten as long as they provide the same functionality to the
outside, their compatibility requirements to previous versions are low.
Firmware (Single Build)
Firmware is code that either provides all functionality for a device (such as the firmware of a
remote control) or that enables a device to load other code (such as the BIOS of an IBM-compatible
computer enable the computer to load its operating system).
Firmware can further be broken into upgradeable and non-upgradable code.
The external requirements for firmware tend to be rigid, but limited in scope, because the
fulfillment of greater requirements, if any, will tend to be left to one or more of the operating
18

systems whose loading the firmware enables. Firmware needs to support specific hardware on one
end and specific functionality on the other. It does not save documents or provide APIs.
The compatibility needs of upgradable firmware are the same as that of drivers.
Table 2: IDPD intensity by category
Type Creates
Documents
Has API Compatibility
needs
Scope IDPD
Non-
Deployable
- - - - -
Firmware - - - - 0
Platform
Software
- + - O 0
Infrastructure
Software
- + +/0 + +
Application
Software
+ O +/0 + +

The IDPD intensity can be explained as follows: non-deployable support software has no
meaningful increments. Firmware and Platform Software don’t need to be compatible to previous
versions, but may build on them. Infrastructure and Application Software are likely to have a major
decline because of the expectations of users described in Table 2 ( - : Very Low, O : Low, + : High).

19

2 Background and Related Work
2.1 Major Cost Estimation Models
2.1.1 Cost Estimation Considerations
Organizations planning sizable incrementally developed projects will be interested in
estimating the cost of those increments. This applies to projects that have their number of
increments or requirements determined and set in advance (like custom-built applications) as well
as to open-ended projects (e.g. operating systems).
Conventional cost estimation methods do not fit incremental development well because they
have no provisions for incremental development. The COCOMO II cost estimation model, for
example, uses cost drivers that remain constant over the course of the whole project. Because of
their staying the same, they are unable to explain the changes in productivity over the course of an
incrementally developed project. Therefore this model in its current form is unsuitable for the
estimation of the cost of incremental development.
Two different points of views for cost predictions are possible regarding incremental
projects and may be combined:
1. Trying to determine the cost of all increments before the beginning of the project, and
2. Trying to determine the cost of the next increment or remaining increments while the project
is ongoing.
The first type allows us to estimate the full cost of a project and therefore to decide whether
to go ahead with it in light of the cost. However, the prediction will be less precise because it cannot
allow for changes in the project that arise during its course.
20

The second type has the advantage of allowing us to take the history of the project into
account and to therefore make more precise predictions. This history cannot be known before the
project starts. On the other hand, by its nature, it does not help in deciding whether it will make
financial sense to go ahead with the project.
In light of the characteristics of these two approaches, it appears to make the most sense to
combine them so a reasonably precise cost estimate can be made before the start of the project, but
the estimate can be refined while the project is ongoing, and action can be taken where necessary in
order to prevent cost getting out of hand.
For an a-priori prognosis of how the productivity of a project will develop, it would be
helpful to find an average percentage factor by which software productivity declines from one
increment to the next (IDPD factor). This factor may vary for different categories of projects.
In order to clearly define the scope of the research, some definitions and categorizations are needed.
They follow in the next several sections of this chapter.
The general types of possible cost estimation models that have been identified (Boehm B. ,
1981) and could be applied to non-incremental development can also be applied to incremental
development:
1. Algorithmic (also known as parametric)
2. Expert Judgment
3. Analogy
4. Parkinson
5. Price-to-Win
6. Top-Down
21

7. Bottom-Up
Out of these, the only one that makes sense to base a verifiable, predictive model of IDPD
cost estimation is the first. This is for the following reasons:
Expert Judgment relies on the presence of experts, who may not always be available. Analogy relies
on the existence of comparable projects. Parkinson (expand the work to fit the budget and
resources) is not project-oriented (and not seriously predictive). Price-to-Win is also neither project-
oriented nor predictive. Top-Down and Bottom-Up are lateral to our model: They can be used to
refine the model further in a given case and to make the prediction more precise, but are themselves
not models.
A parametric model has the advantage that it is predictive, does not depend on the
availability of resources outside the project at hand (save for the ability to set the parameters
reasonably) and that its accuracy and quality can be objectively measured on the basis of existing
project data.
There are several different parametric cost estimation models, with different sets of
parameters that don’t all map to each other. They include (STSC, 2010):
• True S (formerly Price-S)
• SEER for Software
• Sage
• REVIC
• SLIM
While it would be scientifically viable to base an IDPD model on parameters from each of
these models or a synthesis of all of them, the scope of this research dictated that only one be
22

chosen. The choice fell on COCOMO II not because it is of higher quality than the others, but
because of familiarity and the fact that if an IDPD model could be based on COCOMO II
parameters, it should be possible to map it to the other models. Another aspect is that most of the
collected parameters for projects that formed the basis of our case studies were for the COCOMO II
model.
2.1.2 COCOMO II Cost Drivers and Scale Factors
It may be helpful to consider how cost drivers and scale factors behave over the course of an
incremental project when looked at per increment. While cost drivers and scale factors may
certainly develop in different ways depending on the specifics of a given project, it still holds some
interest to attempt to predict their trends over the course of an incremental project that is managed
and executed by rational actors and does not undergo a lot of upheaval. This will at the least allow
an informed discussion about expectations as to whether overall there will be a decrease of
productivity as well as which parameters, be they cost drivers or a scale factors, will most likely
experience it.
Note that an expected trend of “None” doesn’t mean that there should not be any trend, just
that it can go either way depending on the situation of the project.
While Personnel Continuity should not be influenced by the attributes of incremental
development, it is very influential on other projects, which means that a company that wants to be
productive in incremental development should make efforts to retain their workforce, at least during
the course of a given project.

23

Table 3: Effect of COCOMO II effort multipliers on IDPD
Name Description Expected
IDPD trend
Comments
RELY Required reliability Increase May increase for
safety/security critical
systems
DATA How much test data
used per line of code
Increase May increase for big-data
analytics
CPLX Complexity of the
project
Increase May increase with need to
become part of system of
systems or to support
increasing demands
RUSE Instances of reuse Increase Reused components may
need further capabilities
DOCU Amount of
documentation
None Maybe some documentation
growth
TIME Percentage of CPU
time used
None Only relevant to very large
projects; will increase then
STOR Percentage of storage
space used
None Only relevant to very large
projects; will increase then
PVOL Volatility of platform
used for development
Decrease In the beginning, platform
hasn’t solidified yet
ACAP Analyst capability Decrease Will rise if personnel
continuity good
PCAP Programmer capability Decrease Will rise if personnel
continuity good
PCON Personnel continuity Increase High turnover reduces
productivity plus affecting
experience factors
APEX Applications
experience
Decrease Will rise if personnel
continuity good
PLEX Platform experience Decrease Will rise if personnel
continuity good
LTEX Language and tools
experience
Decrease Will rise if personnel
continuity good
TOOL Quality of
development tools
used
Decrease Better tools will be acquired
over time
SITE Collocation and
communication
Decrease If continuity is good, teams
grow closer
SCED Percentage of needed
time available
Increase Schedule compression
becomes more expensive as
system size grows
24

Table 4: Effect of COCOMO II Scale Factors on IDPD
Name Description Expected IDPD
trend
Comments
PREC Precedentedness of the
project
Decrease Project becomes more
familiar
FLEX Procedural flexibility None
RESL Architectural, risk
resolution
Increase Weak architecture
increases rework
TEAM Quality of team
interactions
Decrease If continuity is good,
customer-developer
teams grow closer
PMAT Process maturity Decrease CMMI level increases
should reduce rework

All this means that in general, numerous factors can affect the effort involved in creating a
new increment. As these can vary by increment, they can cause productivity declines and increases
from increment to increment, but overall IDPD effects should be evident.
2.2 Cost Model for Maintenance and Reuse
Maintenance can refer to upkeep effort that has to be expended on
• The codebase,
• Its documentation, and
• Training of personnel and users.
Work on the codebase can consist of
• Adding code,
• Modifying code,
• Deleting code.
25

There are different models for counting the code that is being added, modified and deleted.
One option is of course to fully count the SLOC.
For cases of reuse (for code that is treated as a black box) and adaptation of code (white
box), in some common models (such as COCOMO II (Boehm, et al., 2000)) the effort is calculated
by converting the code size to equivalent SLOC (ESLOC) based on how much effort will have to go
into the integration of that code.
COCOMO II has a nonlinear reuse model that takes into account the understanding of the
software (SU), the degree of assessment and assimilation necessary to integrate fully reused
software modules and the unfamiliarity of the programmers with the software (UNFM).
Additionally, the percentages of the existing software’s design, code and previous integration effort
that would be modified are taken into account as DM, CM and IM parameters.
For maintenance, COCOMO II uses the same model if the amount of the code base changed
is less than or equal to 20%. Above that, a maintenance model is used that takes into account code
size, SU and UNFM. Documentation will have to be updated as needed. Training will have to be
done to the degree needed in order to keep up productivity.
While there is always a basic level of necessary maintenance due to factors such as
technological progress, discovery of bugs and others, there are situations where peaks occur. These
have included Y2K (addressed by fixing the date format and handling of years) and Sarbanes-Oxley
(changes in accounting standards).
Maintenance has also been necessary in many cases in order to adapt to 64-bit addressing
(transition from 32 to 64 bit largely took place in the first decade of the 2000s) and to adapt to
paradigm changes, such as network security (awareness of which has started in the late 1960’s as
26

the Rand Corp., and reached critical mass with the formation of CMU-SEI Computer Emergency
Response Team (CERT) in the late 1980s). Other occasions have included migration from
mainframe to desktop computers (1970s/80s) command line oriented operations to GUI-focused
ones (1980s/90s) and from desktop to mobile computing (ongoing, started in the early 2000s).
2.3 Laws of Software Evolution
The increasing pace of change in software-intensive systems has increasingly shifted
software engineering processes from single-pass, fixed-target development to a multiple-pass,
evolutionary development process. Among other changes, this has changed the scope of software
evolution to include not only relatively small maintenance changes to large operational systems, but
also evolutionary development of similar-sized increments. This raises questions about the
applicability of traditional principles or laws of software evolution to current and emerging forms of
evolutionary development.
Understanding software evolution improves the ability to create processes for effective and
consistent software development. Software evolution is defined as the process of going through the
development, deployment and maintenance of software systems.
2.3.1 Lehman’s System Types
Lehman (Lehman & Belady, 1985) categorizes all software systems to fit one of three types:
S, P or E. This classification was derived from a series of empirical observations Lehman and
Belady made regarding large system development.

27

S-Type Systems
S-type or static-type systems are based on static, formal and verifiable sets of specifications
with easy to understand solutions, such as firmware incorporated in computer hardware. These
simple systems are generally not subject to evolution and therefore Lehman’s laws don’t apply.
P-Type Systems
P-type systems, or practical-type systems, can have their requirements defined precisely and
formally, but the solution is not immediately apparent and its acceptability is dependent on the
environment in which the program is being used. The key difference to S-type is that those are
written to adhere to a static specification, but P-type programs are iteratively developed in order to
improve their effectiveness.
E-Type Systems
The final type of system proposed by Lehman is the E-type, or embedded-type. E-type
programs are defined as all programs that ‘operate in or address a problem or activity of the real
world’. Their evolution is a direct consequence and reflection of ongoing changes in a dynamic real
world. Lehman’s laws only describe the behavior of these projects, which mechanize human or
societal activities and become part of the world they model (Lehman M. , 1980) (Lehman, Ramil,
Wernick, & Perry, 1997).
2.3.2 Lehman’s Laws of Software Evolution
Lehman's laws describe how E-type software projects evolve over their releases. The
original study Lehman and Belady conducted studied IBM software development of OS/360 in the
1970s. After this study, additional papers were written which expanded and amended the set of
laws.
28

For non-trivial incremental development, the earlier increments will exist for some time
before the project is finished (if there even is an end planned – some projects have continued to
evolve). They will therefore either evolve or at least have a considerable age when the later
increments are built.
This applies to some, but not all programs observed here. These laws have been reviewed in
the context of several major open source software projects, with some found to hold and others not
(Xie, Guowu, Chen, & Neamtiu, 2009).
The following is a discussion about the individual Laws of Software Evolution as stated by
Lehman in (Lehman M. , 1980) (Lehman, Ramil, Wernick, & Perry, 1997).
Some notes and caveats for the discussion are:
• The effort attributed to an increment also includes effort spent on previous increments
during the time the new increment is developed.
• Usefulness is regarded as an indispensable part of what makes the quality of a piece of
software sufficient. The authors posit that while other attributes than usefulness are part of
the quality of software, developing a useless piece of software is pointless.
• There is no distinction made or intended between the terms “program” and “system”.
Reflecting the evolution of Software Engineering toward Systems and Software
Engineering, Lehman’s focus has shifted from software to systems over the 17 years that
passed between the publications of (Nguyen, 2010) and (Lehman, Ramil, Wernick, & Perry,
1997), but the laws are intended for both, although the data support is based on software.
A “Reasoning” section is provided when a law’s statement is in need of explanation.
29

2.3.3 Discussion of the individual Laws of Software Evolution
First law: Continuing Change
Statement
A program that is used and that as an implementation of its specification reflects some other
reality, undergoes continual change or becomes progressively less useful. The change or decay
process continues until it is judged more cost effective to replace the system with a recreated
version (Lehman M. , 1980). Note that this “law” does not apply to many embedded software
programs, which usefully control thermostats, carburetors, or most elevators because those
programs tend to have highly stable requirements.
Application to IDPD
Incremental development means that increments build on each other. After each increment,
there is a fully functional and tested program (or system – Lehman talks of “programs” in (Lehman,
Ramil, Wernick, & Perry, 1997) and “systems” in (Lehman, Ramil, Wernick, & Perry, 1997)
reflecting the widening scope of software engineering, the terms will therefore be used
interchangeably for the duration of this discussion).
That the program is used and reflects reality just means that it is an E-type system. The extensions
to systems also excludes S-type systems such as roads, bridges, and piers.
The consequence of this law for the parts of the evolving systems to which it applies is that
from the point in time that any increment gets released, the quality of the code base will decay
unless effort goes into changing or replacing it. The change mentioned in the law can be any type of
maintenance or even design change. It will result in the addition, deletion or modification of code. It
is possible not to put any effort into this part of the existing codebase, but this will result in a
30

reduction of quality and is therefore not a reasonable course of action for a rational actor. While no
effort would be incurred in this case, money would be effectively spent due to the reduction of the
quality of the overall system. Over time, the utility of the system would settle at a low point or
degrade all the way to zero.
Assuming that the effort that can go toward a given increment is largely invariant (as stated
by the Fourth Law further below), this means that a part of the effort for that increment will have to
be directed toward keeping the existing code base useful. Together with the fact that the code base
is growing over time as increments are added over time, this means that the part of the effort
devoted to maintaining the quality of the code base will steadily increase and that the part that can
be used to develop new code will steadily decrease. Even assuming that all other parameters stay
the same (i.e. the productivity as expressed in SLOC per person-hour is the same for later
increments), this will result in a lower new code output per time interval, therefore reducing the
productivity in terms of new SLOC per person-hour.
The only situation in which this effort could be negligible is if the increments are developed
in rapid succession or in cases where the reduction in quality is minimal.
Summary
The productivity in later increments will decline due to the rework and maintenance that has
to be done on the earlier ones. However, little quantitative information has been available, except
for generic information about overall effort breakdowns (Lientz, & Swanson, 1980), domain-
dependent maintenance staffing levels (Boehm B. , 1981), and recalibration of software
maintenance effort estimation models (Nguyen, 2010). The validity of this “law” will be tested by
31

hypotheses 1: i.e., if IDPD=0, then no effort was needed to maintain the earlier increments quality
and value.
Second Law: Increasing Complexity
Statement
As an evolving program is continually changed, its complexity, reflecting deteriorating
structure, increases unless effort is applied to maintain or reduce it (Larman, 2004).
Application to IDPD
As has been stated in the discussion of the First Law, if usefulness is to be retained, the
existing code base will have to be constantly changed (i.e. maintained) as the system evolves. The
Second Law reflects the fact that with all maintenance, there are two layers of effort that need to be
addressed if the quality of the whole system is to be kept equal:
1. The coding work itself
2. Integration of the coding work done into the system in terms of code integration,
documentation, adaptation of the design and rework of its other sections.
The second kind of effort is addressed in this, the Second Law. As with the First Law, there
is the option of not spending the effort, but the outcome will – again – be a loss of quality because
the individual parts of the system will become less and less integrated, creating “rough patches” that
are the results of adding code that is not well integrated, be it to fix deficiencies or to add
functionality.
As in the case of the First Law, increments will require progressively more effort due to this,
reducing the productivity from increment to increment.
32

If no effort is being directed toward reducing the complexity, the complexity will still add to the
effort of the later increments by way of making their development more complex in turn because
the more complex a previous increment becomes, the more complex it will become for the later
increments to integrate with the previous ones.
Summary
Changing previous increments adds complexity, which in turn adds to the effort needed for
the later increments. Again, though, quantification of this effect has been scarce and difficult. Again
the validity of this “law” will be tested by hypotheses 1.
Third Law: Fundamental Law of Program Evolution
Statement
Program evolution is subject to a dynamics which makes the programming process, and
hence measures of global project and system attributes, self-regulating with statistically
determinable trends and invariances (Larman, 2004).
Reasoning
This law is a generalization of the fourth and fifth laws (Larman, 2004). As with the other
laws, laws 3-5 have large major exceptions, such as the effects of Y2K and Sarbanes-Oxley, and
numerous individual exceptions, such as the effects of mergers and acquisitions on corporate
infrastructure software.
Application to IDPD
This means that any change or variance in one system attribute will also be relevant for all
others and that given enough information, the interdependencies of variables in a system should be
predictable.
33

More specifically, the Fourth and Fifth Law state that organizations gravitate toward
stability and that organizational dynamics and the need to maintain familiarity with the system
impose upper limits on the output (that is, if that output is supposed to have any acceptable level of
quality and the growth of the system is to be maintainable).
A fitting metaphor is therefore that of a cover that gets pulled in different directions. For
IDPD, this means that effort spent on one part of the system will directly influence how much effort
can be spent on another. The same applies for different features of the system.
One possible example is that time spent on making the system more secure would take away
from the time spent on improving the user interface. (This is because the law of Conservation of
Organizational Stability says that the work rate per increment is invariant, see below.) Another
interpretation of the Third Law is that the evolution of a program is a predictable process where
similar parameters should yield similar results. This means that the decline in productivity over the
increments – IDPD – should also be predictable.
Summary
The amount of effort to be spent on any given increment is limited and there are
interdependencies between effort parameters. Productivity is often predictable based on parameters
(Nguyen, 2010). The validity of this “law” will be tested by hypotheses 2: i.e. statistical invariance
implies that IDPD will be relatively constant across increments.
Fourth Law: Conservation of Organizational Stability (Invariant Work Rate)
Statement
During the active life of a program the global activity rate in a programming project is
statistically invariant (Larman, 2004).
34

Reasoning
Organizations are striving for stability and there is a point where “resource saturation” is
reached and more resources will not make a difference (Larman, 2004). Frequently, organizational
inertia leaves a maintenance activity with a relatively fixed budget or staffing level over the years.
Application to IDPD
General software engineering experience contradicts at least the literal meaning of this law:
The activity rate can very well vary over the course of a project. This is evident because it is always
possible to reduce the level of activity from the previous one (unless the initial one is already nil,
but then there would be no actual programming project and the law would not apply). The law
therefore needs to be interpreted.
A useful interpretation of this law is that there is an upper limit to the resources that can
reasonably be committed to and the effort that can go into the development of a given increment.
Beyond that limit, the effect of committing more resources or effort will either be little to no benefit
or even detrimental. (This is in line with the well-known insight that “adding manpower to a late
software project makes it later” (Brooks , 1975).
Therefore, according to this law, there is no way for an organization to address the
maintenance and integration effort that needs to go into existing increments by simply expending
more resources in the way of hiring additional programmers. It is inevitable that this effort will take
away from the activities that can be done on the later ones, reducing the productivity achievable for
the new increment.

35

Summary
Beyond a certain upper limit, adding more resources or effort does not benefit the system in
a meaningful way. Again, there is exceptions, such as Y2K. Again, statistical invariance will be
tested by hypotheses 2.
Fifth Law: Conservation of Familiarity (Perceived Complexity)
Statement
As an E-type system evolves all associated with it, developers, sales personnel, users, for
example, must maintain mastery of its content and behavior to achieve satisfactory evolution.
Excessive growth diminishes that mastery. Hence the average incremental growth remains invariant
as the system evolves (Lehman M. , 1980).
Application to IDPD
Everyone working on an incrementally developed project must maintain mastery of its
content and behavior. The bigger the project gets, the more effort will have to be spent on this.
Therefore, since the size of the project will grow over the course of its development, more and more
effort will have to go into this and less will be able to go into new development.
If it is considered that the work rate is invariant, then there is an upper bound to the amount
of effort that can go into any increment. Since the increment to be developed needs to make use of
the existing increments, these will have to be maintained. The law of Conservation of Familiarity
states that there is such a thing as “mental maintenance” which needs to be performed on the minds
of the people working with the existing increments because otherwise, regardless of their quality,
there is nobody who can integrate them.

36

Summary
Even if an increment may have solidified to the point that it needs little or no maintenance
anymore, effort will have to go into the mental states of the people working with it. Excessive
growth may be attempted for an increment, but the mastery of the system will have to keep up with
the increments, so either the growth will not be able to take place or more training will have to be
done after the increment, which reduces that increment’s productivity and evens it out.
Sixth Law: Continuing Growth
Statement
The functional content of E-type systems must be continually increased to maintain user
satisfaction over their lifetime (Lehman M. , 1980).
Application to IDPD
All systems evaluated here are E-type systems. The systems are fully tested and functional
after each increment. The lifetime of each increment runs until the overall system is
decommissioned.
Therefore the functional content of every increment must be continually increased over the
lifetime of the overall system if user satisfaction is to be achieved (which can reasonably be
assumed to be the goal of the development of any system). The effort on this will decrease the
productivity of the current increment because it is done during the same time, and has only the
budget for the current increment.

37

Seventh Law: Declining Quality
Statement
The quality of E-type systems will appear to be declining unless they are rigorously
maintained and adapted to operational environment changes (Lehman M. , 1980).
Application to IDPD
This “law” looks like a weaker phrasing of the one about Continuing Change that adds
nothing of value: It’s not relevant for software development and evolution whether the quality of an
E-type system “appears” to be declining. The decisive question is whether it actually does. If it does
not, then whether or not the systems are maintained and adapted is irrelevant. This law is therefore
only of relevance if that section is changed to “will decline unless”. If it is considered additionally
that E-type systems respond to the environment by definition and that quality and being satisfactory
are the same attributed, then this is the law about Continuing Change restated with different words.
The considerations about its application to IDPD are therefore the same as the ones regarding
Continuing Change.
Hypotheses 1 will test this over the long run. Hypotheses 2 will test “continuality”: there
may be increments that focus more on improving quality rather than functionality.
Eighth Law: Feedback System
Statement
E-type evolution processes constitute multi-level, multi-loop, multi-agent feedback systems
and must be treated as such to achieve significant improvement over any reasonable base (Lehman
M. , 1980).

38

Application to IDPD
A significant amount of parameters will have to be controlled in order to make incremental
development a success. The productivity of a given increment and its quality is relevant to all
following increments.
Lehman E-type systems don’t exist in a vacuum. In order to stay relevant and useful to their
environment, they have to react to the feedback from that environment, which they influence in
turn. Reacting to the environment causes maintenance effort on the existing increments, which
reduces the productivity of the later ones.
This further supports that effort going into the quality of earlier increments will improve the
later increments but also take away productivity from them.
Summary
This “law” just appears to be a restatement of the definition of an E-type system. There may
be different levels of interaction with the surrounding world that increase or decrease the required
feedback work and thus IDPD. This will be tested by hypotheses 3.

39

3 Research Approach
3.1 Behavioral Analysis
In order to arrive at a statistical model of IDPD, the attributes of increments as well as the
parameters of the projects, their personnel and their environments have to be collected. Therefore,
for each increment of projects that had available data, the following has been collected:
• The amount of code added, modified and deleted for each increment in SLOC
• The dates of their increments (where available)
• COCOMO II cost driver and scale factor levels (to the extent that individual drivers and
factors were available)
The collected environmental data (cost drivers, scale factors) is then put in relation to the
quantitative data of the increments themselves (SLOC, dates) and compared via such methods as
correlation and regression analysis.
3.2 Data Collection and Analysis
Due to the diverse nature of incremental software projects, the data collection process
involves establishing and applying consistent criteria for the inclusion of completed incremental
projects. Data has been collected from various sources that fall into the three main categories of
• Software industry
• Controlled experiments
• Open source
40

As with most software engineering research, opportunities for data collection are limited
(Valerdi & Davidz, 2009) due to the size and length of most software engineering projects, and the
fact that many are developed for proprietary or military applications. Knowing this, software
engineering project data was still solicited from industry affiliates supporting this research. Due to
the diverse nature of incremental software projects, the data collection process involves establishing
and applying consistent criteria for inclusion of completed incremental projects.
Only data from projects that satisfy all of the following criteria is collected:
• Starting and ending dates are clear
• Has at least two increments of significant capability that have been integrated with other
software (from other sources) and shown to work in operationally-relevant situations
• Has well-substantiated sizing and effort data by increment
• Less than an order of magnitude difference in size or effort per increment
• Realistic reuse factors for COTS and modified code
• Uniformly code-counted source code
• Effort pertaining just to increment deliverables
As part of data collection process, I interviewed programmers, held workshops and
described the model and core data attributes of the projects such as actual effort, actual size, and
rating levels of the COCOMO II cost drivers. Effort is collected in person-hour and converted into
person-month. Since definition of person-month might differ per organization (by 10% to 20%), the
standard COCOMO model that defines 152 hours per person-month is used (Boehm, et al., 2000).
The size metrics are collected by using code counting tools such as USC’s Unified Code count
41

(UCC)
1
to compare differentials between two baselines of the source program. The size metrics are
based on the logical SLOC definition, adopted from the Software Engineering Institute (Park, 1992)
and adopted into the definition checklist for source statement counts (Boehm, et al., 2000). This
checklist defines an executable statement as one logical SLOC, while non-executable statements
such as blanks and comments are excluded.
Table 5: Core collected data attributes
Metric Description
Product Name of the software product
Build
Build number. A software product that has multiple builds
(increments)
Effort Total time in person-hours for a given increment
New SLOC SLOC count of new modules
Reused SLOC SLOC count of reused modules
COTS SLOC SLOC count of COTS modules
CM The percentage of modified code
DM The percentage of design modified
IM
The percentage of implementation and test needed for the
reused/COTS modules
SU Software Understanding
UNFM Programmer Unfamiliarity
Cost drivers Rating levels for 22 cost drivers

1
http://csse.usc.edu/ucc_wp/
42

3.2.1 Controlled Experiment
In order to have better control over data collection and project parameters, a controlled
experiment was conducted with four simulated application software projects that were given to 21
graduate students of Computer Science. The amount of new requirements and personnel
composition of the projects and related teams were changed throughout the project at irregular
intervals. The number of requirements ranged between zero and more than the teams could
reasonably be expected to finish. In some cases, impossible requirements were given.
It turned out that in the experiment with graduate students it was impossible to reasonably influence
some cost drivers:
• TIME (Execution Time Constraint, the share of available execution time being used)
• STOR (Main Storage Constraint, the share of available storage being used)
• RELY (Required Software Reliability, the impact of a software failure)
• RUSE (Developed for Reusability, the extent to which the software is intended to be reused)
The lack of meaningful CPU and storage limitations in current hardware configurations
made TIME and STOR meaningless for the small projects assigned. Since the subjects knew that
their software was not going to be used in real-world situations, RELY was not going to be a
modifiable influence. Similarly, due to the small scale of the projects, influencing RUSE was not
practical either.
Some personnel-related cost drivers can only change slowly and in most cases within a
limited ranges per person. These include ACAP, PCAP, APEX, PLEX and LTEX (Analyst
Capability, Programmer Capability, Applications Experience, Platform Experience and Language
43

and Tool experience). They therefore had to be changed indirectly by moving personnel between
teams.
Data was collected from the students before the formation of teams, after each week on their
projects and at the end of the semester.
Internal validity considerations concerned mainly the veracity of the submitted time sheets,
while external validity considerations were the small scope of the projects, lack of acceptance
testing, different motivation of the students than that of industry programmers and the influence of
IDE use on code size.
3.2.2 Open Source Projects
Collection of cost estimation parameters for open source projects would have to happen after
the fact since open source organization usually do not collect this data over the course of a project.
In the cases of the open source projects that were used for this research, cost estimation parameters
were not collected. The focus was on observing whether the productivity in new SLOC per person-
month was decreasing. While fluctuations in cost estimation parameters may explain some of the
productivity variations that were observed, it was assumed that over the course of the minor
versions of open source projects that SLOC data was collected here, these parameters would not
vary significantly. The selection of open source projects was based on the stability of projects
(stable release schedule, patch updates, etc.) and their sponsors (i.e. corporate open source). In such
cases, it was assumed stable and constant staff size across increments and the effort was calculated
based on the man-days between the releases.
44

3.2.3 Data Collection Challenges
Data conditioning is a major challenge in the software data collection and analysis process.
Major issues found in the collected data received from commercial sources have included (Boehm
B. , Future Challenges for Systems and Software Cost Estimation, 2009):
1. Inaccurate, inadequate or missing information on modified code (size provided), size change
or growth, average staffing or peak staffing, personnel experience, schedule, and effort
2. Size measured inconsistently (different tools for different increments)
3. Replicated duration (start and end dates) across all increments
4. Low number of increments (less than 3)
5. History of data is unknown
It should be obvious that any one of these issues can severely reduce the quality of the data
to the level of being unsuitable for research. This issue can be addressed through the collection of
data from controlled experiments and open source projects. Table 6, summarizes data sources for
commercial, controlled experiment and open source software:
Table 6: Summary of data sources
Typical
KSLOC
SLOC
collection
Parameter
collection
Main issue
Software
Industry
>100
Various tools,
unverifiable
Unverifiable Secrecy
Controlled
experiments
<5
UCC, source
available
Authors
Individuals have
great influence
Open source Any
UCC, source
available
Authors
Hard to collect
parameters
45

In order to collect cost estimation parameters for open source projects, this needs to happen after the
fact since open source organization usually do not collect this data during the course of a project,
particularly for effort expended by contributors.
3.3 Contextual Analysis
While a fully detailed statistical model of IDPD has been elusive, there may be different
classes of incremental projects that exhibit different patterns or levels of IDPD due to their different
characteristics.
One approach is to separate projects into domains by their position in the hierarchy between
applications, as shown in Table 7, on the top and firmware on the bottom, with some consideration
given to support software (Refer to 1.4.5 detailed explanation of project domains). Envisioned
domain taxonomies include project size, project time frame, functional domain or which part of the
software industry a project is most closely related to.
Table 7: Incremental project domains
Category Description Expected IDPD
Application
Software
Applications that may build on
infrastructure software
+ (High)
Infrastructure
Software
Software that provides services to
other software
++ (Higher)
Platform Software
Drivers that are integrated with
hardware
0 (Medium)
Firmware (Single
Build)
Simple cases of firmware that
cannot be updated
0 (Medium)
Support Software Non-deployable throw-away code n/a
46

3.4 Statistical Analysis
3.4.1 Linear Correlation
Linear correlation is used to measure the strength of the linear association between two
paired sets of data. By computing the correlation coefficient, which is a number between -1 and 1,
the degree of the relationship between two paired sets of data can be determined. The closer to 1 it
is, the more confidence there is of a positive linear correlation and the closer to –1 it is, the more
confidence there is of a negative linear correlation. When the correlation coefficient is zero or close
to zero, there is no evidence of any relationship. Corresponding p-values are looked up in the
standard statistical tables based on the size of data set and its correlation coefficient. The p-value is
a number between 0 and 1, which represents the probability that this data would have arisen if the
null hypothesis were true.
3.4.2 Tailed Pair T-Test
A t-test is used to compare the means from two sets of data in order to test the probability to
accept the null hypothesis. When the two sample sets of data are dependent, a paired t-test is used.
Since the difference between the means is unknown and the first mean can be different from the
second in either direction, a two-tailed t-test is performed
Confidence intervals express the range within which a result for the whole population would occur
for a particular proportion of times an experiment or test was repeated among a sample of the
population. It is computed from the following formula:
!"# = $%"− '"#(,%"+ '"#(+
47

A confidence level refers to the percentage of all possible samples that can be counted as
reliable. For example, if a significance level of 0.05 is used, the 2-tailed test uses half of the
significance level to test the significance in one direction and the other half in the other direction.
That means a significance level of 0.025 in each tail of the normal distribution of the test statistics.
The null hypothesis can be either accepted or rejected based on the significance level of the results
There are two methods that are used to determine the significance of the results. One is to compare
the p-value of the sample with the predetermined significance. If p-value is less than the threshold,
then it indicated strong evidence against the null hypothesis. Otherwise, the null hypothesis cannot
be rejected. The other way is to compare the absolute value of t-stat with confidence value at the n-1
degrees of freedom. Confidence value is computed from a normal distribution with confidence
level, standard deviation and the number of observations in the given sample.
T-stat is computed from the following formula:
_-" =
. −/0
12
√4
5

M is the sample mean and μ0 is the hypothetical mean if the null hypothesis were true. SD is the
standard deviation and N is the number of subjects in the sample.
For a 2-tailed t-test, the null hypothesis can be rejected if your sample t value is more
extreme than the critical t value in either positive or negative direction. If
|_-"| < '"#(, then the null hypothesis cannot be rejected at the given
confidence level, which means that the probability of getting get a confidence value at least as
extreme as the t-stat value is greater than the given confidence level, if the null hypothesis were
true.
48

3.4.3 One-Way ANOVA
One-way ANOVA is a way to test the equality of three or more means at one time by using
variances. In this case, the question that needs to be answered is whether the IDPD of the three
categories differs in a statistically significant way. This can be ascertained using one-way ANOVA,
which is a way to determine whether the means differ significantly. An overall F test is performed
to determine if there is any significant difference existing among any of the means. It is calculated
by the division of between-groups variance and within-groups variance. Between Groups variance
is the explained variance that is due to the independent variable, the difference among the different
categories (referring to the categories in table 7). For example the difference between the overall
IDPD decline in application domain and the overall IDPD decline in platform domain would
represent explained variance. Within Groups variance is the variance within individual groups,
variance that is not due to the independent variable. For example, the difference between the overall
IDPD decline for one project in application domain and another project in the same group would
represent the within groups variance.
F-ratio is the ratio of two sample variances and is computed to determine the p-value. If p-
value is well below the predefined significance level, then it can be concluded that the groups are
statistically significantly different from one another.

49

4 Analysis and Representation of Data
This section summarizes the key data analyses and their results conducted in building and
testing the IDPD hypothesis. Section 4.1 provides an overview of the data selection and
normalization results that defines the baseline data sets for the data analyses. Sections 4.2, 4.3 and
4.4 reports on the data analyzed for productivity decline, build-to-build behavior and domain
hypotheses respectively. Section 4.5 reports on the threats to validity.
4.1 Summary of Data Collection and Selection
Data selection and normalization are completed before most data analyses are started. A set
of project data points, where collected from industry affiliates, open source projects and controlled
experiments. As discussed in Section 3.4, simple and straight forward browsing through data points
helped us eliminate most defective (missing effort, size, or important data definitions such as
counting method or domain information) and duplicated data points.
Error! Not a valid bookmark self-reference. plots a summary of data and the productivity
for each project that was used in this study. Due to the large range of values, data was presented on
a logarithmic scale. Figure 4 provides a histogram showing the distribution of projects by size.

Figure 3. Productivity for all projects on a logarithmic scale
Figure
. Productivity for all projects on a logarithmic scale
Figure 4. Average SOLC of projects
50

Figure 5. Histograms showing number of increments across projects
4.2 Productivity Decline: Hypothesis
The Productivity Decline Hypothesis states, in incrementally developed software projects
that have coherence and dependency between their increments, productivity
course.
Several data points have been analyzed for their IDPD effect and we evaluate several ways
of finding a model that fits our observed productivity declines. The data used for this study are
mostly in the application and infrastr
that fits the data reasonably well while still having some generality and predictive quality. While
some types of trend line can be excluded purely due to their shape not being a good fit in gen
purely logical reasons, others need to be evaluated on goodness of fit metrics.
Although a linear trend line could conceivably turn out to be the best fit over a stretch or
even all the observed data during the course of a project, its predictive
0
5
10
15
20
25
30
35
40
Number of increments
Histograms showing number of increments across projects
Hypothesis 1
Productivity Decline Hypothesis states, in incrementally developed software projects
that have coherence and dependency between their increments, productivity declines over their
Several data points have been analyzed for their IDPD effect and we evaluate several ways
of finding a model that fits our observed productivity declines. The data used for this study are
mostly in the application and infrastructure domains. Our goal is to find an approximation function
that fits the data reasonably well while still having some generality and predictive quality. While
some types of trend line can be excluded purely due to their shape not being a good fit in gen
purely logical reasons, others need to be evaluated on goodness of fit metrics.
Although a linear trend line could conceivably turn out to be the best fit over a stretch or
even all the observed data during the course of a project, its predictive quality is inevitably negated
Number of increments
Number of increments
51
Productivity Decline Hypothesis states, in incrementally developed software projects
declines over their
Several data points have been analyzed for their IDPD effect and we evaluate several ways
of finding a model that fits our observed productivity declines. The data used for this study are
ucture domains. Our goal is to find an approximation function
that fits the data reasonably well while still having some generality and predictive quality. While
some types of trend line can be excluded purely due to their shape not being a good fit in general for
Although a linear trend line could conceivably turn out to be the best fit over a stretch or
quality is inevitably negated

Number of increments
52

when there is a productivity decline. This is because a decreasing trend line will cross the x-axis.
After that point, a linear trend line would therefore predict that productivity is negative, which
wouldn't be a credible prospect for investment in incremental development.
That alone would not necessarily mean that there’s a problem, because one limitation of our
productivity model is that it uses SLOC as a measure of output. In many cases a product can
actually become better when its SLOC are reduced, such as when repetitious code becomes
converted into a subroutine or when code is optimized. Such optimization work is undoubtedly
productive, though it would not be productive as far as the productivity model is concerned.
Therefore it would be possible to have several iterations where SLOC is reduced. However, the
ultimate problem with a linear trend line and its predictive quality would be that when a project
continues long enough, inevitably its overall SLOC would fall below zero SLOC, which is not
feasible.
Another exclusion situation arises with polynomial approximations. Here, with a high
enough order of the polynomial, any number of observations can be approximated fully in a way
that makes the coefficient of determination equal to 1. As with the previous case of the linear line,
the problem is the predictive quality: The polynomial would simply predict that the productivity
would continue to increase or decrease depending on what way the curve went for the last
increment, with possibly grotesquely “wild” results.
4.2.1 Case Study 1: Quality Management Platform
The Quality Management Platform (QMP) is a medium-size commercial software
application that serves to improve the software development process for small and medium-sized
organizations. Functionalities such as process definition, project planning, data reporting,
53

measurement and analysis, defects tracking etc. were incrementally developed as its main features
over six years. More features are planned and scheduled to be added in the future. The increments
were originally regarded as maintenance projects for the first build, but later re-categorized as
incremental development due to showing characteristics of incremental development (Tan, 2009).
Figure 6 shows the productivity over six increments. As tends to be the case in all
productivity graphs so far, a logarithmic trend line (R-squared of 0.53) fits the data better than a
linear one (R-squared of 0.37), though neither fits it particularly well since the R-squared is not very
close to 1. This is because of the increase in productivity in increments 4 and 5, followed by another
decrease in increment 6.
The authors of the study on the project explain the productivity changes in the fourth and
following increment mostly using COCOMO II cost drivers (Tan, 2009) as shown in Table 8.
Table 8: COCOMO II Driver ratings
Build TEAM PCON ACAP PCAP APEX LTEX PLEX
Build 1 NOM HI NOM NOM NOM HI VLO
Build 2 HI HI NOM NOM HI HI LO
Build 3 HI HI NOM NOM HI HI NOM
Build 4 HI HI NOM NOM NOM HI HI
Build 5 VHI LO NOM NOM NOM HI HI
Build 6 VHI LO HI NOM NOM HI HI

The conclusion from this case study of QMP is that while IDPD happens when cost drivers
are largely constant, a large improvement in those drivers can offset IDPD and lead to increases in
productivity, at least over a few increments. Once the improvements have been fully realized, IDPD
will set in again. This is what we’re seeing in the sixth increment here. (Even if we assume that the

cost drivers all improve up to COCOMO II’s optimal values, they can eventually not be improved
upon anymore, so IDPD will set in again.)
Figure 6

4.2.2 Case Study 2: XP1 and XP2
Projects XP 1 and XP 2 were two commercial web
mining developed in Java using a process similar to Extreme Programming (XP). The programmers
were graduate students and detailed logs of activities were kept
and Figure 8 show the productivity over different increments for projects XP1 and XP2
respectively.
Judging by the coefficient of determination, R
here at 0.60, followed by a power one at 0.50 and an exponential one at 0.46.
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
1 2
Productivity in KSLOC/PM
cost drivers all improve up to COCOMO II’s optimal values, they can eventually not be improved
et in again.)
6: QMP productivity across increments
Case Study 2: XP1 and XP2
Projects XP 1 and XP 2 were two commercial web-based client-server systems for data
mining developed in Java using a process similar to Extreme Programming (XP). The programmers
were graduate students and detailed logs of activities were kept (Alshayeb & L
show the productivity over different increments for projects XP1 and XP2
Judging by the coefficient of determination, R-squared, a logarithmic trend line is the best
0.60, followed by a power one at 0.50 and an exponential one at 0.46.
3 4 5 6
Increment
QMP
QMP
Log. (QMP)
Power (QMP)
Expon. (QMP)
Linear (QMP)
54
cost drivers all improve up to COCOMO II’s optimal values, they can eventually not be improved

server systems for data
mining developed in Java using a process similar to Extreme Programming (XP). The programmers
(Alshayeb & Li, 2006). Figure 7
show the productivity over different increments for projects XP1 and XP2
squared, a logarithmic trend line is the best
Log. (QMP)
Power (QMP)
Expon. (QMP)
Linear (QMP)

“No data was available on the number of changes and the effort per change, but the percentage of
total story development effort by story number for Project 1 shows an average increase from one
story to the next of about 6% per story number for refactoring effort and about 4% percent per story
for error fix effort. The corresponding figures for project 2 are 5% an
rates of increase, and while clearly not as good as the anecdotal experiences of agile experts, they
are more likely representative of mainstream XP experience. The small decrease in rates from
Project 1 to Project 2 indicates th
(Bohem & Turner, 2003).
Refactoring and error fixing efforts increased over time. The new design effort seems to be
the dominating effort and the error fixing effort is the smallest of all effor
was negatively correlated with refactoring and error fixing. There does not appear to be any
correlation between refactoring and error fixing efforts
Figure 7
0.00
0.20
0.40
0.60
0.80
1.00
1 2
New Development Effort
“No data was available on the number of changes and the effort per change, but the percentage of
development effort by story number for Project 1 shows an average increase from one
story to the next of about 6% per story number for refactoring effort and about 4% percent per story
for error fix effort. The corresponding figures for project 2 are 5% and 3%. These are nontrivial
rates of increase, and while clearly not as good as the anecdotal experiences of agile experts, they
are more likely representative of mainstream XP experience. The small decrease in rates from
Project 1 to Project 2 indicates there was a small, but not dominant, XP learning curve effect”
Refactoring and error fixing efforts increased over time. The new design effort seems to be
the dominating effort and the error fixing effort is the smallest of all efforts. The new design effort
was negatively correlated with refactoring and error fixing. There does not appear to be any
correlation between refactoring and error fixing efforts (Alshayeb & Li, 2006).
7. XP1 productivity across increments
3 4 5 6
User Story
XP1
Project 1
Log. (Project 1)
Expon. (Project 1)
Power (Project 1)
55
“No data was available on the number of changes and the effort per change, but the percentage of
development effort by story number for Project 1 shows an average increase from one
story to the next of about 6% per story number for refactoring effort and about 4% percent per story
d 3%. These are nontrivial
rates of increase, and while clearly not as good as the anecdotal experiences of agile experts, they
are more likely representative of mainstream XP experience. The small decrease in rates from
ere was a small, but not dominant, XP learning curve effect”
Refactoring and error fixing efforts increased over time. The new design effort seems to be
ts. The new design effort
was negatively correlated with refactoring and error fixing. There does not appear to be any

Log. (Project 1)
Expon. (Project 1)
Power (Project 1)

Figure 8
4.2.3 Statistical Significance
Productivity Decline Hypothesis is
relationship between build-to-build productivity and number of increments.
table 9 indicates a negative correlation between build
increments. If increment increases, productivity falls. In addition, probability values (p <0.1) for
each project show a strong negative relationship between build
increments. Thus these results are compatible with Lehman “laws” 1 on Continuing Change and 2
on Increasing Complexity.
Figure 9 illustrates the normalized productivity of
increments, with 1.00 being the productivity of the initial increment. While the productivity changes
per increment vary greatly and in some cases a lot of productivity can be gained back between
increments, no project has been able to keep its productivity at the level of
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1 2 3
New Development Effort
8. XP 2 Productivity across Increments

Productivity Decline Hypothesis is tested by linear correlation model to assess the
build productivity and number of increments. The result shown in
indicates a negative correlation between build-to-build productivity and number of
increments. If increment increases, productivity falls. In addition, probability values (p <0.1) for
each project show a strong negative relationship between build-to-build productivity and number of
Thus these results are compatible with Lehman “laws” 1 on Continuing Change and 2
illustrates the normalized productivity of eight of our case studies over their
.00 being the productivity of the initial increment. While the productivity changes
per increment vary greatly and in some cases a lot of productivity can be gained back between
increments, no project has been able to keep its productivity at the level of the initial increment. For
3 4 5 6 7
User Story
XP 2
Project 2
Log. (Project 2)
Power (Project 2)
Expon. (Project 2)
56

tested by linear correlation model to assess the
The result shown in
build productivity and number of
increments. If increment increases, productivity falls. In addition, probability values (p <0.1) for
productivity and number of
Thus these results are compatible with Lehman “laws” 1 on Continuing Change and 2
of our case studies over their
.00 being the productivity of the initial increment. While the productivity changes
per increment vary greatly and in some cases a lot of productivity can be gained back between
the initial increment. For
Log. (Project 2)
Power (Project 2)
Expon. (Project 2)
57

any given observation, the loss of productivity (IDPD) from the beginning of the project divided by
the number of increments up to that point ranges from 4% to 91%.

Table 9. Correlation coefficient (with p-value) between build-to-build productivity and number
increments
Project
Names
Correlation
Coefficient
P – value
1 -0.9265 0.0120
2 -0.8243 0.0219
3 -0.7500 0.0430
4 -0.8575 0.0069
5 -0.8675 0.0282
7 -0.9152 0.0425
8 -0.7097 0.0242
Apache 1.3 -0.5813 0.0012
Apache 2.0 -0.6938 0.0002
Perl 5.004 -0.7756 0.0616
Perl 5.005 -0.9473 0.0264
GFS
Kernel
-0.5372 0.0028
php 4.3 -0.7424 0.0110

Figure 10 shows the slopes of the productivity of the case studies between the first and last
increments. The IDPD factors vary between 6.44% and 37.20% per increment, with an arithmetic
average of 16.79% and the median being 14.11%. The range of the slopes for straight lines from the
first to the last increment is narrower: The slopes vary between 5.49% and 15.09%, with an average
of 10.07% and a median of 10.50%.
58

Figure 9. Normalized productivity by increments for 8 projects

Figure 10. Normalized productivity trend lines
It is easy to argue that there is IDPD in projects that have a certain degree of coherence
between increments. One real-world case study in which cost drivers stayed largely constant
showed a near perfect alignment with a trend line using a power formula.

0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 2 3 4 5 6 7 8
Productivity
Normalized Productivity by
Increments
Project 1
Project 2
XP 1
XP 2
QMP
Project 6
Project 7
Project 8
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 2 3 4 5 6 7 8
Productivity
Normalized Productivity Trendlines
Project 1
Project 2
XP 1
XP 2
QMP
Project 6
Project 7
Project 8
59

Perl 5.005

Perl5.17

Project 1

Project 8

Project 2

Php 5.3

Project 3

Php 5.4

60

Figure 11. Overview of productivity trends across projects
Project 4

Project 5

Project 6

Project 7

gfs kernel

61

Another case study shown was not aligned well with a trend line, but had increases with
productivity that could be explained with an improvement in project parameters, such as COCOMO
II cost drivers.
All the projects selected for this study are shown in Figure 11. For each project, the data
points represent normalized productivity value for a specific build. The linear trend line is shown
for each project across different builds.
4.3 Build-to-Build Behavior: Hypothesis 2
Build-to-Build behavior hypothesis states the rate of productivity decline from increment to
increment is relatively constant. Although some projects and “Laws” (Lehman M. , 1980) (Lehman,
Ramil, Wernick, & Perry, 1997) suggest that there is a statistically invariant percentage of
productivity decline across increments, this may not be the case in general.
Build-to-Build behavior Hypothesis was tested by 2-tailed pair t-test model to assess the
relationship between build-to-build slope and overall slope. Results are shown in table 10.
Typically, a null hypothesis states that there is no actual relationship between variables. Since the
hypothesis here declares that IDPD could decline, rise or stay the same, in order to form a
statistically correct test, the hypothesis and its null hypothesis was reversed for the testing purpose.
In other words, the hypothesis is that IDPD is constant across all increments and the null hypothesis
is that the build-to-build IDPD could decline, rise or stay the same. T-tests were performed on the
sample data to examine whether or not the null hypothesis could be rejected.
The significance level α is set to 0.05, which corresponds with a confidence level of 95%.
All the p-values (in Table 10 and Figure 12) are much bigger than the predetermined significance
62

level (p < 0.1). It indicates that the null hypothesis cannot be rejected under the observed result.
Therefore, based on the sample data, it cannot be concluded that the change of productivity from
build-to-build remains constant, which means the productivity could decline, rise or stay the same
from one increment to the next.
Table 10. Confidence Value, t-stat value, and p-value of 2 tailed pair t-test
Project
Name
Confidence Value
(with α α α α = 0.05)
T-stat (absolute
value)
P-value (2 tailed
probability)
1 0.1920 0.1589 0.8835
2 0.2246 0.1514 0.8875
4 0.2459 0.0816 0.9355
5 0.1317 0.0539 0.9578
7 0.1726 0.0229 0.9874
Sendmail
8.14
0.3911 0.2150 0.8384
axc 100 12.2599 0.0147 0.9884
perl 5.004 18.0668 0.2712 0.8038
perl 5.005 27.3963 10.7037 0.9631
perl 5.17 22.4554 0.3767 0.7151
GFS Kernel 9.4272 0.0224 0.9823
php 5.3 13.006 0.0726 0.9428
php 5.4 5.0253 0.2089 0.8377
php 4.3 9.526 0.023 0.9823
php 4.4 3.0079 0.1158 0.9134

Additionally t-stat values and confidence values with n-1 degrees of freedom are computed
for comparison. Since the absolute value of t-stat is less than the confidence value, there is a level of
confidence greater than 95% that the null hypothesis cannot be rejected, and that Lehman “laws” 3
and 4 claiming productivity-decline invariance are not supported by the data.

63

Figure 12. P-value of 2 tailed t-test across projects
4.4 Domain Variation: Hypothesis 3
The domain variation hypothesis states, for different domains (IDPD types), the average
decline in productivity over the course of a project varies significantly. It was tested using the one-
way ANOVA model to test the IDPD percentage difference among the application, infrastructure
and platform domains. In this research, the IDPD percentage difference among the application,
infrastructure and platform domains were examined.
As shown in Figure 13, the average of overall IDPD decline is 18% for Infrastructure
domain, 9% for application domain, and 5% for platform domain. The distribution of IDPD across
software domains is shown in Figure 14 and Figure 15.

0
0.2
0.4
0.6
0.8
1
1.2
Poject 1
Project 2
Project 4
Project 5
Project 7
Sendmail 8.14
axc 100
perl 5.004
perl 5.005
perl 5.17
GFS Kernel
php 5.3
php 5.4
php 4.3
php 4.4
Projects
P-Value

Figure 13: Averages of Average IDPD
Figure

Table 11. One
Table 10. Sum of
squares
Between
Groups
0.0589
Within
Groups
0.0673
Total 0.1262
One-way ANOVA was used to test for
11 shows the output of the one
9%
18%
0%
5%
10%
15%
20%
IDPD
Domain Names
0%
5%
10%
15%
20%
25%
30%
35%
0 2
Application

: Averages of Average IDPD
Figure 14: IDPD Distribution over Domains
Figure 15: Average IDPD over Domains
. One-way ANOVA results for Hypothesis 3
Sum of
squares
df Mean
Square
F Sig
0.0589

2 0.0294

8.7453

0.002

0.0673 20

0.0034

0.1262 22

way ANOVA was used to test for average IDPD differences among three domains.
shows the output of the one-way ANOVA analysis and whether we have a statistically
5%
Domain Names
5
0
6
3
6
0 0
2
0
0
1
2
3
4
5
6
7
Number of Projects
Domain
4 6 8
Application Infrastructure platform
64

Distribution over Domains

ces among three domains. Table
way ANOVA analysis and whether we have a statistically
0
0%-10%
11%-20%
21%-30%
10
65

significant difference among our group means. Average IDPD for different domains differed
significantly across the three sizes, F (2, 20) = 8.7453, p = .002 (p < 0.05). The significance level is
0.002, which is below 0.05. Therefore, there is a statistically significant difference in the mean
average IDPD among different domains. Taken together, these results suggest that overall IDPD
differs across the three different domains. At this level, though, the results extend Lehman “law” 8
in identifying domains with differing IDPD percentages.
4.5 COCOMO II and IDPD
4.5.1 Cost Estimation
Organizations planning sizable incrementally developed projects will be interested in
estimating the cost of those increments. This applies to projects that have their number of
increments or requirements determined and set in advance (like custom-built applications) as well
as to open-ended projects (e.g. operating systems).
The types of possible cost estimation models that have been identified and could be applied
to non-incremental development can also be applied to incremental development include
Algorithmic (also known as parametric), Expert Judgment, Analogy, Parkinson, Price-to-Win, Top-
Down and Bottom-Up.
Out of these, a parametric model has the advantage that it is predictive, does not depend on
the availability of resources outside the project (save for the ability to set the parameters reasonably)
and that its accuracy and quality can be objectively measured on the basis of existing project data.
Though not all parameters of all models can be mapped or converted to each other, the large
majority of them can be (Madachy & Boehm, 2008). For this reason, it becomes feasible to evaluate
data collected using any model with any other model.
66

Table 12. COCOMO II Cost Drivers and Scale Factors expected to have a decrease in effort
Name Description Comments/Special cases
PREC
Project
precedence
Project becomes more familiar
TEAM
Team
interactions
If continuity is good, customer-developer teams
grow closer
PMAT
Process
maturity
CMMI level increases should reduce rework
RUSE
Instances of
reuse
Reused components may need further
capabilities
Later increments will tend to see less reuse
with the exception of systems that are built top
to bottom
DOCU
Amount of
documentation
Foundational increments need the most and
best documentation.
PVOL
Development
platform
volatility
In the beginning, platform has not solidified yet
ACAP
Analyst
capability
Analysts will gain capabilities if personnel
continuity good
PCAP
Programmer
capability
Programmers will gain capabilities if personnel
continuity good
TOOL
Quality of
development
tools
Better tools will be acquired over time
SITE
Collocation and
communication
If continuity good, teams grow closer and
communications improve

While it would therefore be scientifically viable to base an IDPD model on parameters from
each of these models or a synthesis of all of them, the scope of this document dictated that only one
be chosen. The choice fell on COCOMO II not because it is of higher quality than the others, but
because data collected using its parameters was most readily available and the fact that if an IDPD
model could be based on COCOMO II parameters, it should be possible to map it to the other
models.
67

It may be helpful to consider how cost drivers and scale factors behave over the course of an
incremental project. Collected industry data suggests cost drivers to be variable across increments.
In contrast to this, the incremental development model for COCOMO II cost estimation model
assumes that the cost drivers remain constant across the increments of the whole project (Boehm &
Lane, 2010).
Constant cost drivers are unable to explain the changes in productivity over the course of an
incrementally developed project. Therefore, this model in its current form is not equipped for the
estimation of the cost of incremental development.
4.6 IDPD and COCOMO II
Adapting COCOMO II to incremental development in order to predict IDPD also presents
an opportunity to re-evaluate some parameters for the current time. While some parameters have
held up over time, some may not have a major influence anymore nowadays (TIME and STOR
come to mind, defined in Table 13).
4.6.1 COCOMO II cost drivers and scale factors
Table 13 and Table 14 list expectations of whether the effort factor related to a given
COCOMO II cost driver is expected to increase, decrease or have no expectation of any trend. The
reason why the tables look at the expectation for effort multipliers instead of cost driver or scale
factor levels is in order to make the tables more readable for readers unfamiliar with the individual
drivers or factors: While for some of those a higher level can mean more effort, for some others it
means less effort. Note that an expectation of no specific effort does not mean that there should not
be any trend, just that it can go either way depending on the situation of the project.
68

Table 13. COCOMO II Cost Drivers and Scale Factors expected to have no specific trend in effort
Name Description Comments/Special cases
FLEX Procedural flexibility
RESL
Architectural risk
resolution
Architecture may require increasing amounts
of rework; but familiarity with architecture
grows
RELY Required reliability
May increase for safety/security critical
systems
DATA
Test data per line of
code
May increase for big-data analytics
CPLX Complexity
May increase with need to become part of
system of systems or to support increasing
demands
TIME
Percentage of CPU time
used
Only relevant to projects with finite CPU
resources or that share their execution time
with others, such as in a mainframe
environment; will increase then, but such
projects are exceptions nowadays
STOR
Percentage of storage
space used
Only relevant to projects with hard upper
limits on storage or that share their storage
with others; will increase then, but such
projects are exceptions nowadays
PCON Personnel continuity
High turnover reduces productivity because
new team members need to be trained and
affects experience related factors. Therefore,
this is a pivotal cost driver for IDPD which
depends on how the project is managed.
APEX Applications experience
Applications experience will increase if
personnel continuity good
PLEX Platform experience
Platform experience will increase if personnel
continuity good
LTEX
Language and tools
experience
Language and tools experience will increase
if personnel continuity good
SCED
Percentage of needed
time available
Schedule compression becomes more
expensive as system size grows
69

That the expectation for all effort factors is to go in no specific direction or to decrease
suggests that in a general case, a well-managed organization which is able to keep turnover under
control should experience a gradual decrease in the effort that is expressed in the existing
COCOMO II cost drivers when applied to individual increments. This is because the organization
will grow more familiar with the project and the developers will become more experienced.
This means that in the case of a well-managed organization and in non-exceptional situations, IDPD
cannot be captured or predicted using existing COCOMO II drivers applied to increments.
Table 14. P-value of weighted multiple regressions between cost drivers and productivity
Project
Drivers
DR-1 DR-2 DR-3 DR-4
Product Factors 0.961 0.945 0.986 0.001
Platform Factors 0 0.789 0.762 0.139
Personnel Factors 0.025 0.029 0.137 0.169
Project Factors 0.136 0.425 0 0.02
Scale Factors 0.249 0.42 0.651 0.014

While Personnel Continuity (PCON) is a parameter that cannot be directly set by an
organization, it is very influential on other parameters. This means that an organization that wants to
be productive in iterative and incremental development should make efforts to retain its workforce,
at least as far as the team of a given project is concerned.
In order to see which cost factors may have a higher influence on productivity and to
increase the accuracy of predicting the IDPD factors in the next build, the relationship between
individual cost drivers and productivity was evaluated by examining the p‐value of regression
between cost drivers and productivity.
70

Since there are 22 cost drivers and scale factors in total, only the ones with p-values within
the significant range were reported (< 0.08) ( see Table 15). In these 4 directed research (DR)
projects, TIME is the factor that predicts the productivity the best. Then we have STOR, PCAP and
SITE (Collocation,), followed by ACAP, RELY, PREC, TEAM and PMAT. A weighted regression
was conducted on all the cost drivers versus productivity (see Table 16).
Table 15. P-value (<0.08) of regression between cost drivers and productivity for 4 DR (directed
research) projects

The results are still not consistent throughout the four DR projects. APEX seems to be the
factor that predicts the productivity the best among all the factors. The ratings of APEX represent
the level of applications experience of the project team developing the software system. Multiple
regressions between groups of cost drivers and productivity, weighted by increment sizes, were also
performed (Table 14). In order to find better predictions for productivity, the added lines of codes
were used as the weight of the regression. The results show that personnel factors and project
factors seem to be the best factors to predict productivity within the 4 projects. Weighted regression
between cost drivers and productivity (only for cost drivers with correlation > 0.7)
Project
Driver
DR-1 DR-2 DR-3 DR-4
TIME 0.0022 0.08 0.08 /
STOR 0 0.04 / /
ACAP / 0.06 / /
PCAP / 0.06 / /
RELY / / / 0.07
SITE
(Collocation)
0.08 0.07 / /
PREC / 0.01 / /
TEAM 0.0031 / / /
PMAT 0.05 / / /
71

Table 16. Weighted regression between cost drivers and productivity (only for cost drivers with
correlation> 0.7)
Project
Driver
DR1 DR2 DR3 DR4
PREC 0.05 0.37 0.93 0.01
CPLX 0.34 0.55 0.94 0.45
RUSE 0.02 0.25 0.10 0.94
TIME 0.38 0.47 0.61 0.72
STOR 0.00 0.94 0.29 0.02
LTEX 0.00 0.62 0.76 0.70
SITE(Communication) 0.40 0.94 0.00 0.79

4.6.2 How IDPD influences parameters
In a project whose increment schedules have been planned beforehand, the size of IDPD will
influence the SCED driver in subsequent increments in cases where one increment takes longer than
expected. This can cause a chain reaction of schedule compressions if adjustments to the schedule
aren’t made.
In situations where team members are judged by their productivity, a loss in productivity can
cause an increase in personnel fluctuation (captured by PCON), which will weigh negatively on
other cost drivers as pointed out further above.
4.6.3 Potential New Cost Estimation Parameters
Some of the main differences between monolithic and incremental projects are: 1) The
existence of previous increments that need to be maintained, 2) The passing of time between
releases, 3) That a large body of code (when compared to the current increment) needs to be
integrated with the current increment, 4) that the previous increments may have a number of known
72

defects, 5) that those increments may be more or less complex, and 6) that later increments may
depend on previous ones to a lower or higher degree.
The first of these is trivially true from the second increment on with regard to the existence
of previous ones. It has been established that previous increments need to be maintained (Moazeni,
Link, & Boehm, 2013). The remaining attributes of individual increments do not correspond with
COCOMO II cost drivers. Since they are all measurable, albeit to different degrees, they hold
interest as potential cost drivers.
Time between increments
The time passing between two increments causes the following phenomena (Moazeni, Link,
& Boehm, 2013):
• Quality/usefulness of the software will slowly decrease all the way to zero because a system
at rest will not keep pace with technological advances.
• Knowledge of the existing increments diminishes by personnel fluctuation or details being
forgotten in the organization.
• New “must have” features need to be added to existing increments due to customer or
market requests.
• Rework may need to be done due to changing paradigms (e.g. desktop to mobile).
Since all of these factors are exasperated the more time passes, more time between
increments should cause the IDPD between them to increase. Increment productivity in terms of
new SLOC over effort from several open-source projects has been found to be aligned with the time
difference to the previous increment in a statistically significant way.
73

Defect number and complexity in increments
Defects of the previous increments will need to be fixed in those increments or compensated
for in newer ones.
While defect complexity may not be easy to measure, it should be possible to at least categorize
them into three categories:
1. Easy (e.g. typo or likely off-by-one)
2. Intermediate (requires fixing some complex interactions, but root cause is known)
3. Difficult (e.g. unknown root causes make system display occasional off-nominal behavior)
Comparative size of current increment
Most projects that data has been collected on have shown a decline in productivity between
the first and last increment. The body of code will generally rise over the course of an incremental
project. It is therefore of interest to relate the amount of existing code to the productivity of a new
increment.
Intuition would say that the more code is pre-existing, the more work will have to be done
on maintenance and integration. In addition to that, a growing body of code will either become
increasingly complex, which causes the effort to master it to grow or work to become necessary to
reduce the complexity (Moazeni, Link, & Boehm, 2013).
Increment productivity in terms of new SLOC over effort from several open-source projects
has been found to be aligned with the size of the previous increment in a statistically significant way
(Table 17).
74

Table 17. Correlation of Code Size Metrics to Increments

Coherence of the project
An influence on the complexity of increments is to what extent these increments need to
make calls to functions of other increments.
The measures here could lie in the number of calls to other increments, the complexity of
these calls, the amount of required data flow over time between increments and the complexity of
the increments that are being called (cyclomatic complexity and other measures).
Evaluated new Cost Drivers
Four potential parameters that would extend COCOMO II for increments have been
evaluated. These were RCPLX (Complexity of Increment I+1 relative to complexity of existing
software), IPCON (Staff turnover from previous increment’s staff), PRELY ( Previous increments’
required reliability) and IRESL (Architecture and risk resolution (RESL) rating of increment I+1
relative to RESL of 1..I).
Initial data collected by the software industry did not show any good correlation with IDPD.
First to current Current to last Previous to current Size per increment
Project Correlation P-value Correlation P-value Correlation P-value Correlation P-value
common-daemon -0.67 0.0086 0.17 0.2987 0.08 0.4024 -0.16 0.3008
tomcat 6.0 -0.465 0.0245 0.4663 0.0245 -0.2806 0.1302 -0.4195 0.0326
php-5.4 0.1028 0.3726 0.0384 0.4484 0.107 0.3603 -0.0068 0.4859
sendmail 8.14 0.5548 0.1684 0.5614 0.1631 -0.1356 0.4112 0.5181 0.1158
php-5.3 -0.6464 0.0003 0.698 0.0001 -0.1264 0.2724 -0.7045 0.0001
gsf-kernel -0.5906 0.0015 0.5905 0.0015 0.1345 0.2772 -0.6184 0.0005
cmirror 0.1 -0.3215 0.1225 0.2063 0.2263 0.3132 0.1304 -0.2302 0.1873
cmirror 2.6.9 -0.0395 0.4111 0.0073 0.4776 0.2264 0.0953 -0.0405 0.4084
cluster 3.1 0.3863 0.2223 -0.7339 0.0498 -0.8014 0.028 0.7578 0.0143
75

Ten projects have been evaluated using measures of effort and size. The best ANOVA p-values
have been found when relating new lines to productivity.
4.7 Threats to Validity
4.7.1 Internal validity considerations
First increment peculiarities
The productivity of the first increment of any software project can be atypical due to a
number of reasons. These include that exploration may occur and that IDEs (Integrated
Development Environments) being used may generate a significant amount of the code that is only
edited or configured in later increments. Similarly, newly added source code may be based on
templates the development team is reusing from other projects. Two further reasons for lower
incremental productivity are that it is doing much of the requirements engineering and architecting
for the later increments; and that it may end with a good deal of partially-developed code not
counted in the incremental release (this may also be the case for later releases). While these are
concerns, the productivity of all examined projects still decreases between the second and last
increments, though to a smaller degree.
Time reporting
Some of our projects had university students in their development teams. The accuracy of
time logs by the students is somewhat questionable. Depending on the amount of credits they are
aiming for, students have to work a certain amount of hours per week. In cases where the students
did not actually work as long as expected, there is a temptation to overstate the time so that it
appears they are doing the work that is expected of them.
76

Another aspect is that some students may be late in filling out the time sheets, and, when
reminded of them, fill them out with poorly-remembered hours. The threat to the internal validity is
somewhat mitigated by the interest in embellishing their hours being common to all student projects
and all parts of the projects, so that statements about whether productivity is increasing or
decreasing remain valid. Similar concerns can apply to professional programmers occasionally.
The accuracy of the time logs as submitted by members of development teams can be
questionable when team members work on several projects or several increments (testing increment
1, developing increment 2, designing increment 3) at the same time and need to attribute parts of
their time to different projects or increments. The threat is mitigated for professional and student
developers by the likelihood of distortions being common to all parts of the project, which will not
significantly affect the study of the development of their productivity.
4.7.2 External validity considerations
Student projects have relevant threats to their external validity due to the motivation of the
students. The motivation of the students is different than that of people working in the software
industry in that the students are typically facing less existential risks for failure. A professional
programmer may get laid off for bad performance and face significant material losses. A student
may face a bad grade in one course.
Students in project courses are generally working part-time, and may have productivity peaks
and valleys due to exams or big homework assignments in other courses. This can also happen to
professional developers working on multiple projects.
77

5 Incremental Development Cost Estimation
Several major cost estimation models are used throughout the software industry. Some are
open to the public while others are closed and can only be used as a commercial product. If they
would address incremental development, this would obviously have an impact on this research.
What follows is Table 18 with the names of the models and a sourced statement on whether they
have provisions for incremental development.
Table 18. Comparison of cost estimation models
Model Incremental development provisions
COCOMO II (Incremental COCOMO) Yes (Boehm, et al., 2000)
SEER for Software No
PRICE H unknown
SLIM No
SAGE unknown

Conventional cost estimation methods do not fit incremental development well because they
have no provisions for incremental development. COCOMO II model has an extension called
Incremental COCOMO that has provisions for Incremental development. In the following section
the Incremental COCOMO II model is described followed by the estimation steps that are part of
the Incremental COCOMO model . This research invalidates the Incremental COCOMO extension
in its current form, therefore, the Constructive Incremental Cost Model (COINCOMO) and its tool
will be introduced.
78

5.1 Incremental COCOMO II
Incremental COCOMO was defined as an alternative to the traditional Waterfall model of
the software development process. This model deals with cost estimation of software projects with a
series of concurrent increments yielding an intermediate product. Increments have the following
MBASE/RUP phases:
• Product Design (PD)
• Detailed Design (DD)
• Code and Unit Test (CUT)
• Integration and Test (IT)
At a high level, the estimation for incremental development involves:
• Use COCOMO II model with scale factor and cost driver ratings
• The percentage of change in software size from all previous increments (Adaptation
Adjustment Modifier, AAM)
• The starting milestone for each increment after the first
• Any delay time for starting an increment
• The distribution of effort and schedule for each phase
5.1.1 Incremental COCOMO Estimation Steps
1. Compute the size for each increment and the total size
Adjust the size of the succeeding increments after the first increment due to reusing the
software product from all the previous increments.
79

9:.<=
>
= <=
>
+ ?99@ × B<=
C
>DE
CFE
G
AAM is applied to the size of all previous increments. AAM is based on the percent design
modified (DM), percent code modified (CM), the integration and test modified (IM), the
assessment and assimilation (AA), software understanding (SU) and programmer
unfamiliarity (UNFM).
99H = I0.4 ×L@M+ I0.3 ×@M+ I0.3 ×@M
99@ =
$99+99HO1+OI0.02 ×R ×RSH@MTU
100
99H ≤ 50
99@ =
$99+99H +IR×RSH@M+
100
, 99H ≥ 50
2. Determine the total effort and schedule for the total single product
Calculate total effort and schedule using the COCOMO II model and distribute the estimates
for MBASE/RUP phases.
3. Determine increment 1’s estimated effort and schedule
Use the COCOMO II effort calculation formula and calculate the effort using COCOMO II
parameters, size and phase distribution.
4. Determine the added integration effort
The added integration effort is calculated by subtracting the full product effort and the sum
of estimated effort for all of the independently estimated increments.
5. Distribute the added integration effort across remaining increments
6. Determine all remaining increments’ effort and schedule
80

Use COCOMO II model to estimate the total effort for the remaining increments. Use the
increments’ adjusted size.
7. Distribute each increments’ effort and schedule over the phases
5.1.2 Incremental Development Estimation Considerations
The COCOMO II cost estimation model, uses cost drivers that remain constant over the
course of the whole project. Because of their staying the same, they are unable to explain the
changes in productivity over the course of an incrementally developed project. Therefore this model
in its current form is unsuitable for the estimation of the cost of incremental development.
5.2 Constructive Incremental Cost Model
Build-to-Build behavior hypothesis states that the rate of productivity decline from
increment to increment is not constant which means that productivity may decline, rise or stay the
same. In a multi-build incremental development model, this means that COCOMO cost drivers
(Effort Multipliers and Scale Factors) should be variable across increments. This contradicts with
the incremental development model in Appendix B of the COCOMO II book (Boehm, et al., 2000)
which assumes that the cost drivers remain constant across the increments. To overcome this, a new
tool called COINCOMO (COnstructive INcremental COst MOdel) is developed to enable the cost
drivers to be variable across increments.
5.2.1 Software Engineering Reasons for Multi-Build, Overlapping RUP Cycles
The selection of multiple builds with overlapping MBASE/RUP cycles was done for
sensible systems/software engineering reasons. The primary one for multiple builds was based on
productivity in the face of risks, in particular for unprecedented System of Systems. An all or
81

nothing approach to the software development (also known as a "big-bang" approach) does not
make sense. With multiple builds, the program has a chance to re-scope either pieces of software or
whole systems at the decision points based on Feasibility Evidence. It also provides the ability to
generate and demonstrate (in an earth-based test environment) some of the system capabilities
earlier than is possible in a big-bang approach, and those capabilities should involve many of the
"risky" bits, like communications with the constellation of spacecraft modules and network
management of the constellation.
Employing overlapping, concurrent builds for the different systems was also selected for
sensible systems/software engineering reasons. Assuming there is not a software development
resource restriction, and that developers have appropriate "application" domain experience, then
getting the teams to work concurrently is the smartest thing to do from an overall software
development effort perspective: the efforts of the teams are additive and not based on the total
Source Lines-of-Code (SLOC) count being used on any given system. From a systems and
software engineering perspective, producing builds that are fully systems-tested, even if parts of the
build may use simulation and/or emulation, means that both the software-to-system-integration
teams and systems-test teams can start to work earlier than in a big-bang approach. The systems
engineers and software architects can practice the concurrent "Agile Architecting and Stabilized
Development" approach, for systems with high concerns for safety or security or requirements that
are unstable or emerging. The early integration and test experience will also increase the acquirer’s
confidence about the capabilities and architecture of the system.

82

5.3 Incremental Development Cost Estimation using COINCOMO
The original COCOMO I model used the waterfall model for its anchor points: the cost and
schedule estimates covered the effort between Software Requirements Review (SRR) and Initial
[software] Operational Capability (IOC). That model, which dates back to the 1970s (Royce, 1970)
has been recognized as an appropriate subset of the possible process models under the appropriate
conditions) (Barry & Port, 1999) . Current terminology for anchor points, as used in COCOMO II,
requires a more general approach (Bohem B. , 1996).

Figure 16. Overlap across Builds

83

The co- evolution of the RUP in the late 1990s led to the obvious alignment of the anchor
points (and their assessments) as gatekeepers to the RUP “phases”. The gatekeepers are Inception
Readiness Review (IRR) for Inception phase; Life Cycle Objectives (LCO) for Elaboration; Life
Cycle Architecture (LCA) for Construction; and IOC, with a Transition Readiness Review (TRR),
for Transition; and a Product Release Review (PRR) to formalize the end of the ”Transition”. These
anchor points, super-imposed on the RUP activities diagram, are shown in Figure 17. Also shown in
Figure 17 via shading is the area from LCO to IOC, i.e., the Elaboration and Construction phases,
and the covered activities that are included in the total effort that a COCOMO II estimation covers.
Builds made using the COCOMO II model follow either the “once-through” or “incremental”
development strategies, as described in Appendix B of the IEEE’s J-STD-016-1995 (IEEE Standard
Association, 1995). In both cases, there is only one system integration and software qualification

Figure 17. RUP with MBASE Anchor Point

84

test, which are two of the primary activities taking place during the Transition phase. These two
development strategies are in contrast to the “evolutionary” strategy that develops a system in
builds, where the first build incorporates part of the planned capabilities, the next build adds more
capabilities, and so on, until the system is complete. If the builds follow the “evolutionary” strategy,
including their integration with the system to be deployed, they are often fully tested and delivered
to the acquirer, in which case all the phases of RUP are repeated for each build. However, how the
phases overlap is not depicted in RUP; Figure 16 shows the possible overlaps across the builds.

Figure 18. Overlap across builds with carry forward

85

5.3.1 COINCOMO Cost Estimation Steps
Bellow are different steps in estimation:
1. Compute schedule and effort for each Increment
Use the COCOMO model to calculate schedule and effort for each build. COCOMO
calculates effort and schedule for the Elaboration and Construction phases of a build with new code
and code carried forward (Figure 18) from the previous build treated as re-used code with re-use
parameters.
Recall from COCOMO reuse model that reuse is influenced by Adaptation Adjustment
Modifier (AAM) which is based on % Design Modified (DM), % Code Modified (CM), %
Integration and test Modified (IM), assessment and assimilation (AA), Software Understanding
(SU) and programmer unfamiliarity with software (UNFM).
99H = I0.4 ×L@M+ I0.3 ×@M+ I0.3 ×@M
99@ =
$99+99HO1+OI0.02 ×R ×RSH@MTU
100
99H ≤ 50
99@ =
$99+99H +IR×RSH@M+
100
, 99H ≥ 50
To calculate the size for the next increment, the sum of previous build’s increments are
treated as reused and adjusted with AAM. This is then carried forward from the previous build and
added with the increment to be developed.
The amounts and re-use factors of the carried code change with the number of times the
code is carried: these factors influence the "Equivalent SLOC", or ESLOC, for the carried code
which in-turn impacts the effort and schedule.
2. Distribute the effort and schedule across phases
86

Use the Constructive Phased Schedule and Effort Model (COPSEMO) to distribute effort
and schedule across the inception, elaboration, construction and transition phases. The COPSEMO
model, separates the man- power loading between Elaboration and Construction, and adds
additional effort and schedule for the Inception and Transition phases.
3. Determine resulting cumulative effort and schedule for each increment
The total effort is additive, meaning it is the sum of effort across all the phases of the
increments. When calculating the total schedule in this approach, only the parts up to an overlap are
counted.

87

6 The COINCOMO Tool
COINCOMO (COnstructive INcremental COst MOdel) is a screen-oriented, interactive
software package that assists in budgetary planning and schedule estimation of a software
development project. Through the flexibility of COINCOMO, a software project manager (or team
leader) can develop a model (or multiple models) of projects in order to identify potential problems
in resources, personnel, budgets, and schedules both before and while the potential software
package is being developed.
The current version of COINCOMO, version 2.0, is a multi-platform software cost
estimation application, which in one tool supports both the cost estimation models of COCOMO
and COPSEMO (Constructive Phase and Schedule Model). Future iterations will be extended to
include other cost estimation models, and will be extensible to support such models.
The tool will allow users to do cost estimation of systems with multiple subsystems, where each
subsystem could have multiple builds. The tool will also support estimation of individual and
multiple builds of a single subsystem; where estimates across builds take into account concurrent
development of overlapping MBASE/RUP phases (Boehm, et al., 2000):
• Inception
• Elaboration
• Construction
• Transition
The full COCOMO II model includes three stages.
1. Supports estimation of prototyping or applications composition efforts.
88

2. Supports estimation in the Early Design stage of a project, when less is known about the
project’s cost drivers.
3. Supports estimation in the Post-Architecture stage of a project. COINCOMO 2.0
implements stage 3 formulas to estimate the effort, schedule, and cost required to develop a
software product. It also provides the breakdown of effort and schedule into software life-
cycle phases and activities from both the Waterfall model and the MBASE/RUP Model
(Boehm, et al., 2000).
The COINCOMO 2.0 implements stage 3 formulas to estimate the effort, schedule, and cost
required to develop a software product. It also provides the breakdown of effort and schedule into
software life-cycle phases and activities from both the Waterfall model and the MBASE/RUP
Model. The MBASE/RUP model is fully described in Software Cost Estimation with COCOMO II.
6.1 COINCOMO Terminology
Before launching into the discussion of COINCOMO, one must understand the terminology
used by COINCOMO. A System is the basic aggregate unit used by COINCOMO to represent a
real-life software development project. A single part of the system is represented by a Sub System
in COINCOMO. Thus, a single and independent functional part of a real-life system may be
represented by a sub-system. Sub-systems are further broken down into Components. For example,
different builds of the same part of a project can be represented in COINCOMO as components.
Components are further broken down into Subcomponents. These are the basic atomic units used by
COINCOMO to describe a software development activity. In a real-life scenario, for one build, new
code development and maintenance of previously developed code for a given component could
qualify as two separate sub components.
89

6.2 The COINCOMO Architecture
In order to meet resource contention and portability capability goals and requirements, a
client-server separation of the User Interface (UI), import/export, and format translation from the
model calculations and database access logic was used. A Java based client UI and JDBC-
compatible server-based RDBMS was used to ensure portability among the required platforms. The
RDBMS-centric design inherently supports extensibility, as methods for extending 3NF schemas
are well known. In addition, the design supports multiple open-source database vendors for
flexibility. An XML-based formatting system allows translation of estimation data to and from
multiple formats, as well as direct data transfer between databases.
The element relationship diagram in Figure 19 shows the primary entities involved with the
COINCOMO 2.0 system. The user interacts with the system by providing inputs based on the use of
multiple cost models. The cost models are database-driven and interact with the users input via the
database. Outputs are put back to the database. Then, the outputs are retrieved and displayed back
to the user.

90

Figure 19: COINCOMO Element Relationship diagram
Figure 20 accommodates for the new System, Subsystem, Component (on the left side) and
multiple subcomponents (on the right side). The COCOMO estimation is shown at the bottom of the
figure.
As shown in Figure 21, the COPSEMO calculation shows the effort and schedule
distribution across the MBASE/RUP phases along with a graphical representation of phased
schedule and effort. The spinners shown in Figure 21, allows adjusting the effort and schedule
percentage for the MBASE/RUP phases.

91

Figure 20: COINCOMO Component estimation

Figure 21: COPSEMO calculation

92

The COINCOMO tool support cost estimation of multibuild software, Figure 16, and allow
adjusting the overlapping across the builds.

Figure 22: COINCOMO support for multibuild estimation

93

7 Conclusion and Future Work
7.1 General Conclusions
The research presented in this dissertation, defined Incremental Development Productivity
Decline (IDPD), formulated hypothesis (Decline hypothesis, Build-to-Build behavior hypothesis
and Domain hypothesis) and based the industry, open source and controlled experiment data,
confirmed or rejected the hypothesis. In summary, most of the analyzed data points are well
aligned with negative trend lines. The average arithmetic decline is about 16.79 and the slopes vary
between 5.49% and 15.09%, with an average of 10.07% and a median of 10.50%. Average IDPD
across domains are 18% for infrastructure domain, 9% for application and 5% for platform software
domain. Finally, the build-to build behavior hypothesis confirmed IDPD to be different across the
builds in a software projects. This hypothesis disproved Lehman Laws 3 and 4 on statistical
invariance.
After evaluating how different kinds of trend functions (linear, logarithmic or exponential)
fit the observed cases of IDPD, it was concluded that there is not one single type of negative trend
function that fits best in all cases, but that different cases had a better correlation with different
types of trend functions. However, we found that on average the trend lines were best aligned with a
logarithmic trend function. This is in line with an IDPD factor determining the ratio of productivity
between subsequent increments, and that factor having a limited variance.
Challenges included that there was no way to retroactively validate and verify the collected
data and that backstories of the projects were often either nonexistent outright or not deep.
94

Additionally, the collected metrics were not always the same and the code counting methods
differed between different projects.
As part of this research, individual COCOMO II parameters were evaluated with regard to
their development over the course of increments and how they influence IDPD. The reverse is also
done. In light of data collected in the controlled experiment projects and the analysis that was
presented in this dissertation, the following conclusions are made:
• Even when conventional COCOMO II drivers are applied to the increments of such projects and
made variable over them, they are unable to explain the cases where a well-managed
organization (meaning one which manages personnel education and turnover well) experiences
IDPD. Therefore, either new parameters need to complement or replace the existing ones in the
model or the prediction has to take place outside the model.
• A controlled experiment with students showed inconclusive results.
• Promising new parameters with statistically significant predictive capabilities in several open-
source projects include the size of the current increment and the time between increments and
the time between the current and the first increment as well as the time between current and the
last increment.
• Data should be collected using parameters that include defect and complexity analysis of code.
7.2 Summary of Contributions
The main contributions of this research are as follows:
• Confirmed nontrivial existence of IDPD (Hypothesis 1)
• Rejected (confirmed null hypothesis) build-to-build constancy of IDPD (Hypothesis 2).
Lehman “Laws” 3 and 4 on statistical invariance not confirmed
95

• Confirmed IDPD variation by domain (Hypothesis 3)
• Developed COINCOMO cost estimation model supporting cost driver variation by
increment
This dissertation described the COINCOMO cost estimation model and its tool for
incremental development estimation. In a study that was done on existing cost estimation models,
the majority of them turned out to be not supporting incremental development. The COCOMO cost
estimation model that was selected as the model for this research supported an incremental model
through an extension. However, as described in this dissertation, the model was invalidated based
on the results of hypothesis 2.
In addition, the study of COCOMO II parameters and IDPD discussed how existing
parameters can be applied to IDPD and proposed new cost parameters.
7.3 Future Work
A much desired outcome of this research would be to construct a mathematical model for
incremental development cost prediction that, for a given project, takes into account:
• Its IDPD domain,
• The COCOMO II (or other major cost estimation model) cost driver ratings applicable for
its individual increments,
• New parametric cost drivers specific to IDPD,
• Other factors to be determined.
Once such a cost estimation model for IDPD has been found and sufficiently verified, it may be
used to extend major cost estimation models such as COCOMO II.
To achieve a mathematical model, future work should focus on employing the following methods:
96

• Collected data needs to be statistically evaluated further under more aspects in order to increase
chances of finding patterns that predict productivity.
• Project data needs to be collected and evaluated regarding the parameters proposed in this paper.
• Additional potential parameters should be evaluated.

97

BIBLIOGRAPHY
Albrecht, A. (1979). Measuring Application Development Productivity. Joint SHARE/GUIDE/IBM
Application Development Symposium, (pp. 83-92).

Albrecht, A., & Gaffney, J. E. (1983). Software Function, Source Lines of Code, and Development
Effort Prediction: A Software Science Validation. IEEE Transactions on Software Engineering ,
SE-9 (6), 639 - 648.

Alshayeb, M., & Li, W. (2006). An empirical study of relationships among extreme programming
engineering activities. Information and Software Technology , 48 (11), 1068–1072.

Elssamdisy, A., & Schalliol, G. (2002). Recognizing and responding to bad smells in extreme
programming. 24th International conference on Software Engineering. ACM.

Boehm, B., & Port. (1999). Escaping the Software Tar Pit: Model Clashes and How to Avoid Them.
ACM SIGSOFT Software Engineering Notes , 24 (1), 36-48.

Lientz, B., & Swanson, E. B. (1980). Software maintenance management: a study of the
maintenance of computer applications software in 487 data processing organizations." Reading,
MA (1980). Reading, MA, USA: Addison-Wesley.

Boehm, B. & Royce. (1989). Ada COCOMO and the Ada Process Model. TRW SPACE AND
DEFENSE SECTOR REDONDO BEACH CA.

Boehm, B. (2009). Future Challenges for Systems and Software Cost Estimation. 42nd Annual DoD
Cost Analysis Symposium .

Boehm, B. (1981). Software Engineering Economics. Englewood Cliffs, NJ, Prentice-Hall.

Boehm, B. W., Clark, Horowitz, Brown, Reifer, Chulani, et al. (2000). Software Cost Estimation
with Cocomo II. NJ, USA: Prentice Hall PTR Upper Saddle River.

Boehm, B., & Lane, J. A. (2010). DoD Systems Engineering and Management Implications for
Evolutionary Acquisition of Major Defense Systems. DoD Systems Engineering Research Center.

Boehm, B., & Port, D. (1999). Escaping the Software Tar Pit: Model Clashes and How to Avoid
Them. Software Engineering Notes, Association for Computing Machinery, (pp. 36-48).

Boehm, B., & Royce, W. (1989). Ada COCOMO and the Ada process model.

Boehm, B. (1996). Anchoring the software process. IEEE Software , 13 (4), 73-82.

Boehm, B., & Turner, R. (2003). Balancing agility and discipline: A guide for the perplexed.
Addison-Wesley Professional.
98

Brooks, & Frederick, P. (1975). The mythical man-month (Vol. 1995). Reading, MA, USA:
Addison-Wesley.

Elssamadisy, A., & Schalliol, G. (2002). Recognizing and responding to bad smells in extreme
programming. 24th International conference on Software Engineering. ACM.

Gill, G., & Iansiti, M. (1994). Microsoft Corporation: Office Business Unit. Harvard Business
School Case Study.

Hurley, D. (1995). Software Engineering and Knowledge Engineering: Trends for the Next Decade.
World Scientific.

IEEE Standard Association. (1995). Trial-use standard for information technology software life
cycle processes software development acquirer-supplier agreement.
Larman, C. (2004). Agile and iterative development: a manager's guide. Addison-Wesley
Professional.

Lehman, M. M., & Belady, L. A. (1985). Program evolution: processes of software change.
Academic Press.

Lehman, M. (1980). Programs, life cycles, and laws of software evolution. 68, no.9, pp. 1060-1076.
IEEE.

Lehman, M., Ramil, J. F., Wernick, P. D., & Perry, D. E. (1997). Metrics and laws of software
evolution-the nineties view. 4th International Software Metrics Symposium (pp. 20-32). IEEE.

Madachy, R., & Boehm, B. (2008). Comparative Analysis of COCOMO II, SEER-SEM and True-S
Software Cost Models . University of Southern California Center for Systems and Software
Engineering .

Madachy, R., Boehm, B., Clark, B., Tan, T., & Rosa, W. (2011). Future Software Sizing Metrics
and Estimation Challenges. 15th Annual Practical Systems and Software Measurement (PSM)
Users' Group Conference,.

Moazeni, R., Link, D., & Boehm, B. (2013). Incremental development productivity decline. 9th
International Conference on Predictive Models in Software Engineering. ACM.
Moazeni, R., Link, D., & Boehm, B. Lehman’s laws and the productivity of increments:
Implications for productivity. 20th Asia-Pacific Software Engineering Conference (ABSEC) (p.
2013). Bangkok, Thailand: IEEE.

Nguyen, V. (2010). Improved Size and Effort Estimation Models for Software Maintenance. PhD
Dissertation, University of Southern California, .

99

Park, R. E. (1992). Software Size Measurement: A Framework for Counting Source Statements.
CARNEGIE-MELLON UNIV PITTSBURGH PA SOFTWARE ENGINEERING INST.

Scacchi, W. (1991). Understanding software productivity: towards a knowledge-based approach.
International Journal of Software Engineering and Knowledge Engineering , 1 (3).

SEBoK. (n.d.). Retrieved from Guide to the Systems Engineering Body of Knowledge :
http://www.sebokwiki.org/wiki/Guide_to_the_Systems_Engineering_Body_of_Knowledge_(SEBo
K)
STSC, S. T. (2010, October). Software Development Cost Estimating Guidebook. Retrieved from
http://www.stsc.hill.af.mil/consulting/sw_estimation/softwareguidebook2010.pdf

Tan, T. (2009). Productivity trends in incremental and iterative software development. 3rd
International Symposium on Empirical Software Engineering and Measurement. IEEE Computer
Society.

Valerdi, R., & Davidz, H. L. (2009). Empirical research in systems engineering: challenges and
opportunities of a new frontier. Systems Engineering , 12 (2), 169-181.
Vosburgh, J. (1984). Productivity factors and programming environments. 7th international
conference on Software engineering. IEEE Press.

Walston, C., & Felix, C. (1977). A Method of Programming Measurement and Estimation. IBM
Systems Journal , 16 (1), 54 - 73.

Xie, Guowu, Chen, J., & Neamtiu, I. (2009). Towards a better understanding of software evolution:
An empirical study on open source software. IEEE International Conference on Software
Maintenance.

COPSEMO, Constructive Phase and Schedule Model, Retrieved from:
http://csse.usc.edu/csse/research/COPSEMO/

Appendix A: The COINCOMO User Interface
COINCOMO 2.0 is a standalone software system intended for a single user. The software is
user- interactive in that it attempts to interface well with a user’s needs, using extensive mouse
interaction wherever possible. In order to use CO
familiar with the user interface (UI) of COINCOMO.
Figure A- 1 shows the initial screen
to the database.
Figure A- 1
As depicted, the COINCOMO UI is divided into three major areas:
Appendix A: The COINCOMO User Interface
alone software system intended for a single user. The software is
that it attempts to interface well with a user’s needs, using extensive mouse
In order to use COINCOMO efficiently, the user
user interface (UI) of COINCOMO.
shows the initial screen presented by COINCOMO to a user once he/she has connected
1. Initial screen after connecting to database
As depicted, the COINCOMO UI is divided into three major areas:
100
alone software system intended for a single user. The software is
that it attempts to interface well with a user’s needs, using extensive mouse
the user must become
presented by COINCOMO to a user once he/she has connected

101

1) Main Menu Bar – This area contains the menu selection of the main functions of
COINCOMO. These selections are File, Parameters, and Help. The menu selections File and
Parameters are discussed in sections 4 and 5 of this manual respectively. The menu selection
Help is used to display the version and copyright information about COINCOMO, and open
up HTML version of this User Manual.
2) Project Browser – This area is used to view and access a project. COINCOMO 2.0 can
display only one project (system) at a time. The Project Browser displays the entire project,
along with all of its sub-systems and components, in a hierarchical or tree fashion. An
example of such a display is shown in Figure A-2 .The Project Browser also contains, at the
top, a Project Toolbar with image buttons that allow the addition or removal of sub-systems
and components to a project. Note that these context-sensitive actions can also be invoked
directly in the Project Browser by right-clicking on a given unit and selecting the
appropriate option from the resulting pop-up menu. These context-sensitive pop-up menus
also allow the user to rename a unit. Two examples of this behavior are shown in A-3.
3) Unit Detail Area – This area shows detailed information about a particular unit (System,
Sub-System, or Component) selected in the Project Browser. There are two tabs in this area:
a. Overview – displays relevant information about the unit selected in the Project
Browser; shows the Summary Report for a system (refer A-4), Phase-Effort-
Schedule for a sub-system (refer Figure A-5) and the Component Level Estimating
Form (CLEF) for a component (refer Figure A-6), respectively
b. COPSEMO – displays COPSEMO calculations for a component (refer Figure A-7);
this tab is disabled for systems and sub-systems.

Figure A- 4

Figure A- 3

Figure A- 5

Figure A- 2
102

103

Figure A- 6

Figure A- 7

We now examine the Component Level Estimating Form (CLEF) in more detail, since this
is where most of the user interaction takes place. A typical CLEF display for a component is shown
in Figure A-8.
Figure A- 8
The important parts of the CLEF, as marked in Figure A-8, are discussed below:
104

1) Add Subcomponent – This button is used to add a sub component to the currently selected
component in the Project Browser.
2) Delete Subcomponent – This button is used to remove a sub component to the currently
selected component in the Project Browser.
3) Aggregate Subcomponent Selection – This drop-down menu can be used to select all,
deselect all or toggle the selection of the various sub components for the currently selected
component in the Project Browser (refer Figure A-9).

4) Scale Factors – This button displays the Scale Factor Dialog Box as shown in Figure A-10.
Figure A- 9
Figure A- 10

Dialog boxes such as these are common throughout COINCOMO 2.0. To make changes to a
given factor, the user needs to click
menu, displaying the values available for the factor (refer Figure
use the Increment button, which also displays a pop
and 75% (refer Figure A-11). After making the desired adjustments, the user should click the
button to propagate the changes to the currently selected component or sub co
factors to their original values, one can click on the
should click on Close.
5) Schedule – This button displays the

Dialog boxes such as these are common throughout COINCOMO 2.0. To make changes to a
given factor, the user needs to click on the corresponding Rating button. This displays a pop
menu, displaying the values available for the factor (refer Figure A-11). To adjust further, one can
button, which also displays a pop-up menu with preset options of 0%, 25%, 50%,
11). After making the desired adjustments, the user should click the
button to propagate the changes to the currently selected component or sub component. To reset all
factors to their original values, one can click on the Reset button. To exit the dialog box, the user
ton displays the Schedule Dialog Box as shown in Figure
Figure A- 11
105
Dialog boxes such as these are common throughout COINCOMO 2.0. To make changes to a
button. This displays a pop-up
11). To adjust further, one can
up menu with preset options of 0%, 25%, 50%,
11). After making the desired adjustments, the user should click the Apply
mponent. To reset all
button. To exit the dialog box, the user
in Figure A-12.
106

Figure A- 12

6) Subcomponent Selection Column – This column is reserved for identifying and selecting a
sub component. Selection is denoted by a tick mark (√) that appears in this column. Only
one sub component can be selected at a time using this column.
7) Subcomponent Name Column – This column is used to house the name of each sub
component located in the CLEF. Clicking on the name of a sub component displays an input
dialog box as shown in Figure A-13. This dialog box can be used to rename a newly added
or existing sub component. To propagate the change in the sub component, the user should
enter the desired name and click on OK. By default, the addition of a new sub component
creates a sub component with the name (SubComponentX), where X is an auto-incremented
number.

Figure A- 13
8) Subcomponent Size Column – This column is used to house the SLOC of each sub
component located in the CLEF. Clicking in this column corresponding to a sub component
spawns a dialog box as shown in Figure A-14. The value for SLOC can be computed in one
of three ways. One way is to enter the value directly in the SLOC field as shown in Figure
A-14. Another way is to use the function point model as shown in Figure A-15. Finally, the
107

Adaptation Adjustment Factor method can be used for the computation of the SLOC as
shown in Figure A-16. The language of implementation of each sub component is initially
unspecified, but may be set using this dialog box. Upon completion of SLOC sizing input,
click on Apply and then Close.

Figure A- 14

108

Figure A- 15
109

Figure A- 16

Figure A- 17
9) Labor Rate Column - This column specifies the amount of money (per month) which a
developer working on a given sub component would be paid. The labor rate for a particular
sub component can be edited by clicking on its corresponding Labor Rate column. This
110

spawns a dialog box as shown in Figure A-18. One can enter the new labor rate here, and
click on OK to affect the change. The range on labor rate is between $0 and $99,999.

Figure A- 18

10) Subcomponent Effort Adjustment Factor (EAF) Column – This column displays the cost
drivers for a specific sub component. By clicking on this field, a dialog box appears as
shown in Figure 3-19.

Figure A- 19
As shown in Figure A-19, the cost drivers are divided into five groups: Product, Platform,
Personnel, Project, and User. By default, the ratings of all cost drivers are “NOM”, and their
111

percentage increments (also referred to as inter-cost driver values) are set at 0%. The user can
manipulate the value of each of these cost drivers individually by changing the corresponding rating
and % increment values. The final rating of a cost driver is calculated using this formula for the
interpolation.
Final rating = (Next cost driver rating - Current cost driver rating) * Current inter-cost driver / 100
As individual cost driver ratings are changed, the total product of the cost drivers also
changes. When all cost drivers have been modified, one can click on Apply and then Close to view
the final EAF, which is displayed in the EAF column.
COINCOMO currently supports only the Post Architecture model of software development.
The Post Architecture model applies once the software architecture has been formulated. This is in
contrast to the Early Design model, which is supposed to be used at the earliest phase of a software
project. In terms of the COCOMO paradigm used by the COCOMO II suite of estimation tools, the
Early Design model differs from the Post Architecture model in its use of Effort Adjustment
Factors. The Early Design model considers only seven pre-defined effort adjustment factors,
whereas the Post Architecture Model makes use of seventeen pre-defined effort adjustment factors,
sixteen of which are shown in Figure A-19.
Estimation Results Area
This area displays the effort, schedule, cost and staff estimates calculated by COINCOMO
for a given component. These statistics are: the total SLOC count for the entire component (SLOC),
total hours per month (PM/M), the total effort in person-months (PM), the total schedule for project
completion in months (M), the total productivity (PROD), the total project cost (COST), the total
project cost per instruction (INST), the total staff requirement to complete the project (Staff), and
the risk associated with the project (Risk).
112

Status Bar
This area at the bottom of the CLEF displays various status and error messages to the user.
Subcomponent Implementation Language Column
This column indicates the development language for a given sub component. By default, no
language is selected for a sub component. This can be changed while sizing the sub component, as
shown in Figure A-16, where a drop-down box for implementation languages is present.
Subcomponent Nominal Development Effort (NOM Effort DEV) Column
This column holds the most likely effort estimate for a given sub component without
incorporating the Effort Adjustment Factors (EAF).
Subcomponent Estimated Development Effort (EST Effort DEV) Column
This column holds the most likely effort estimate for a given sub component obtained by
incorporating the Effort Adjustment Factor (EAF).
Subcomponent Productivity (PROD) Column
This column contains the calculated result of the sub component’s individual SLOC divided
by the sub component’s most likely effort estimate.
Subcomponent Cost (COST) Column
This column contains the most likely estimate of the development cost for a particular sub
component.
Subcomponent Instruction Cost (INST COST) Column
This column contains the most likely cost per instruction for a given sub component. This
number is calculated from the Cost/SLOC for each sub component.
Subcomponent Staff Requirement (Staff) Column
113

This column houses the most likely estimate for the number of full-time software developers
(FSWP) that would be needed to complete a given sub component in the estimated development
time.
Appendix B: Working with COINCOMO
This section discusses the main menu options of COINCOMO 2.0 in more detail.
COINCOMO 2.0 is available in three editions: Desktop Edition, Database Edition, and Unified
Edition. Depending on the edition you obtained, some menu options might not be
available/accessible to you.
COINCOMO 2.0 has two modes of operation: Desktop Mode and Database Mode. Desktop
Edition only operates under Desktop Mode, and Database Edition only operates under Database
Mode; only Unified Edition allows you to switch between the two modes.
File Menu
The expanded File menu from COINCOMO 2.0 is shown in Figure B-1 for Database
Edition and Figure B-2 for Desktop Edition
114

Figure B- 1

Figure B- 2
Connect (Database Mode Only)
This menu selection is used to connect and retrieve project data from the COINCOMO 2.0
database pertaining to a user. To connect to the database, one must click on Connect. A login screen
then appears, as demonstrated in previous section, and also shown in Figure B-3. The user should
enter his/her credentials to connect to the database. Once the user is connected to the database, the
message “Connected to Database” appears on the Status Bar at the bottom of the screen.
COINCOMO 2.0 requires a running database server to operate for Database Edition or in Database
Mode for Unified Edition. Therefore, this selection will result in an error if the database server is
115

not running, or if there is a communication failure between the COINCOMO tool and the database
system.

Figure B- 3
Disconnect (Database Mode Only)
This menu selection is used to disconnect from the COINCOMO 2.0 database. To
disconnect from the database, one must simply click on Disconnect. Once the user is disconnected
from the database, the message “Disconnected from Database” appears on the Status Bar at the
bottom of the screen. Before disconnecting from the database, the current state of the project will be
saved in the database, and the next time the user retrieves the project, this state will be restored.
Hence, adequate care must be taken before disconnecting from the database so as not to save the
current project with improper data values.
116

New Project
This menu selection is used to create a new project. By default, the newly created project is
named “(SystemX)”, where X is an auto-incremented number. It can be renamed using context
menus in the Project Browser.
View Projects (Database Mode Only)
This menu selection is used to view the list of existing projects in the COINCOMO 2.0
database. Clicking on this selection brings up the dialog box shown in Figure B-4. From this dialog
box, the user can choose the desired project, click on Load and then on Close to display the contents
of that project in the Project Browser.
This same dialog box also contains the option to delete a project. Again, one simply needs to
choose a project and click on Delete in order to delete a project. Note that deleting a project will
permanently delete a project and all its contents from the COINCOMO 2.0 database. Hence,
extreme caution is advised while using this option, unless the database system is configured (with
regular backups) to overcome the problems caused by an unwanted deletion.
117

Figure B- 4

Save Project (Desktop Mode Only)
This menu selection will allow you to save the current project to a file in COINCOMO 2.0’s
own file format for later editing, as shown in Figure B-5. Once you save the current project to a file,
any subsequent Save menu selection will also save to the same file. If your current project is opened
through a COINCOMO 2.0 file, this menu selection will also save to that file.
118

Save Project As
This menu selection allows you to save the current project in COINCOMO 2.0’s own file
format in a different file name locally. For Database Mode, this menu selection allows you to
actually save the current project to a local file so you can do offline editing in Desktop Mode later,
and then use the local file to synchronize with COINCOMO 2.0 database.
Export As CSV/HTML/XML
Although COINCOMO 2.0 stores all the project-related estimates and calculated values in
either its own file format or in COINCOMO 2.0 databases, in the real world, such estimates lend
themselves to further processing and presentation if they are in commonly used file formats.
Figure B- 5
119

COINCOMO 2.0 has a data export feature, which allows it to generate such files. This feature is
available through the Export menu selection is shown in Figure B-6.

Figure B- 6
As shown, data related to an entire project can be exported in CSV (Comma-Separated
Values), HTML or XML format. It must be noted that this export feature operates on the current
project. Once a particular export format is selected, a save dialog box pops up to allow the user to
save the files in a desired location. An example of this behavior is shown in Figure B-7.

120

Figure B- 7
While the choice of HTML and XML formats creates only one file per project, an export in
the CSV format creates two different files, named “(SystemX).csv” and “(SystemX)’s Multibuild
Report.csv”, where (SystemX) is the file name you specified. While the first file deals with the
complete estimates for the entire project, the second deals with estimates in relation to multiple
builds within the project, if there are any.
Exit:
This menu selection can be used to exit from COINCOMO 2.0. COINCOMO 2.0 in
Database Mode automatically save any changes to the database immediately; however, to ensure
data integrity, it is recommended that users first disconnect from the database using the Disconnect
option, and then exit COINCOMO using the Exit option. For COINCOMO 2.0 in Desktop Mode,
this menu selection will detect if your current project has been saved or not, and prompt you for the
option to save if necessary.
121

Parameters Menu
The expanded Parameters menu from COINCOMO 2.0 is shown in Figure B-8. This menu
is available when an available component has been selected. Options within this menu allow control
over the model used by COINCOMO 2.0 to generate the effort, schedule and cost estimates.

Figure B- 8

Effort Adjustment Factors
This menu selection presents a dialog box that allows the user to manipulate the default
multiplier values of the various cost drivers used by COINCOMO for a particular rating. Screens
with sample values are shown in Figure B-9 through Figure B-13. To affect a change, one must
simply change a multiplier value, and click on Apply. To set all ratings to their default multiplier
values, the Reset button can be used.

122

Figure B- 9

Figure B- 10

123

Figure B- 11

Figure B- 12
124

Figure B- 13
Scale Factors
This menu selection presents a dialog box (refer to Figure B-14) that allows the user to
manipulate the default values of the various scale factors used by COINCOMO. To affect a change,
one can change a value, and click on Apply. To set all ratings to their default multiplier values, the
Reset button can be used.

125

Figure B- 14
Equation Editor
This menu selection allows the user to change the basic effort and schedule equations used
by COINCOMO 2.0. The menu selection spawns the dialog box shown in Figure B-15, which
allows modification of these equations. To make the changes effective, the user must click on Apply
after making the necessary changes. Reset can be used to set all ratings to their default values.

126

Figure B- 15
Function Points
This menu selection spawns the dialog box shown in Figure B-16. It allows the user to
modify the default values of the Function Point Factors, which are used to calculate the SLOC size
if the Function Point method is used. Again, clicking on Apply will cause any changes to be
effective, while clicking on Reset set all values to default.

127

Figure B- 16

Person Month
This menu selection spawns the dialog box shown in Figure B-17, and allows the user to
enter a new value for the hours included in one person-month. The default value for Person Month
is 152, which means that there are 152 hours in one person-month. To change this value, the user
can enter the new value in the dialog box (e.g., 160), and click on OK. The valid range of values is
between 120 and 184 only.

Figure B- 17

128

Mode Menu
For COINCOMO 2.0 Unified Edition, This expanded menu is available to allow the user to
switch between Desktop Mode (Desktop Edition functionality) and Database Mode (Database
Edition functionality). If you are running in Desktop Mode (by default) and want to switch to
Database Mode or vice-versa, a message will pop up asking if you want to switch to the other mode
if you have project(s) opened currently. Confirming the switch will result in closing of the
projects(s).

Figure B- 18

Abstract (if available)

Abstract Software production is on the critical path of increasingly many program abilities to deliver effective operational capabilities. This is due to the number, complexity, independence, interdependence, and software‐intensiveness of their success‐critical components and interfaces. The estimation parameters and knowledge bases of current software estimation tools are generally good for stable, standalone, single increment development. However, they do not fully account for the degrees of program and software dynamism, incrementality, coordination, complexity, and integration. These phenomena tend to decrease software productivity relative to the cost model estimates made for the individual software components and for the overall systems, but it is difficult to estimate by how much. ❧ Incremental software development generally involves either adding, modifying, or deleting parts of the code in the previous increments. This means that if a useful system is to be built, the maintenance that will have to go into previous increments will take away productivity from the later ones. ❧ This research tests hypotheses about a phenomenon called Incremental Development Productivity Decline (IDPD) that may be more or less present in incremental software projects of various categories. ❧ Incremental models are now being used by many organizations in order to reduce development risks while trying to deliver releases of the product on time. It has become the most common method of software development with characteristics that influence the productivity of projects. ❧ Different ways of measuring productivity are presented and evaluated in order to come to a definition or set of definitions that is suitable to these categories of projects. Data from several sources has been collected and analyzed, and hypotheses tested about the degree of IDPD and its variation by increment and category. The results indicated the existence of an IDPD phenomenon, that its magnitude varies by application category, but that it tended to vary from increment to increment.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

The incremental commitment spiral model process patterns for rapid-fielding projects

PDF

A model for estimating cross-project multitasking overhead in software development projects

PDF

The effects of required security on software development effort

PDF

A model for estimating schedule acceleration in agile software development projects

PDF

Domain-based effort distribution model for software cost estimation

PDF

Composable risk-driven processes for developing software systems from commercial-off-the-shelf (COTS) products

PDF

WikiWinWin: a Wiki-based collaboration framework for rapid requirements negotiations

PDF

Software architecture recovery using text classification -- recover and RELAX

PDF

A reference architecture for integrated self‐adaptive software environments

PDF

A value-based theory of software engineering

PDF

Security functional requirements analysis for developing secure software

PDF

A unified framework for studying architectural decay of software systems

PDF

Toward better understanding and improving user-developer communications on mobile app stores

PDF

Calibrating COCOMO® II for functional size metrics

PDF

A search-based approach for technical debt prioritization

PDF

Using metrics of scattering to assess software quality

PDF

A framework for intelligent assessment and resolution of commercial-off-the-shelf product incompatibilities

PDF

Calculating architectural reliability via modeling and analysis

PDF

Experimental and analytical comparison between pair development and software development with Fagan's inspection

PDF

Quantifying the impact of requirements volatility on systems engineering effort

Asset Metadata

Creator Moazeni, Ramin R. (author)

Core Title Incremental development productivity decline

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Publication Date 08/07/2014

Defense Date 04/14/2014

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag incremental development,OAI-PMH Harvest,productivity decline,software engineering,software measurement

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Boehm, Barry W. (committee chair), Nakano, Aiichiro (committee member), Settles, F. Stan (committee member)

Creator Email moazeni@usc.edu,r_moazeni@yahoo.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-456158

Unique identifier UC11287212

Identifier etd-MoazeniRam-2788.pdf (filename),usctheses-c3-456158 (legacy record id)

Legacy Identifier etd-MoazeniRam-2788.pdf

Dmrecord 456158

Document Type Dissertation

Format application/pdf (imt)

Rights Moazeni, Ramin R.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA