1 product-line architectures don batory department of computer sciences university of texas at...
Post on 28-Dec-2015
215 Views
Preview:
TRANSCRIPT
1
Product-Line Architectures
Don BatoryDepartment of Computer SciencesUniversity of Texas at Austin
2
Introduction
In 1500s, accepted “truth” that the Earth was the center of the Universe
Geocentricity was obvious and largely accurate lunar eclipses positions of “fixed” stars but planetary motion caused problems...
3
Retrograde Motions...
complex models (spheres inside spheres) ultimately failed to predict planetary positions accurately
4
A Revolution
In 1543, Copernicus proposed a radically different explanation of our universe
• heliocentricity elegantly explains retrogrades, forms basis of today’s understanding of planetary systems
(Extreme) example of how science evolves
• negating commonly held “truths” yields models of the universe that not only are consistent with known facts, but are more powerful and lead to deeper understandings
• more common examples lead to incremental advances
5
Universe of Software
class foo { int a; int b; …}
class bar { …}
...
Classes!
Today we elegantly express software as hierarchies of classes and webs of interconnected objects
Global varsand functions
int global;
int func( ) { ... }
void main( ) { global = func( ); …}
...
6
Looking Ahead...
What software design and construction technologies lie in the future?
How will we understand software? What will be our “unit” of
encapsulation? How will we produce & specify
software?
try negating obvious “truths” and see if a consistent explanation of software remains
7
Some Changeable “Truths”
Today
one-of-a-kind applications
design expressed in objects and classes
code our implementations
product-line architectures PLAs
designs expressed in components
codeless programming(software plug-and-play)
Tomorrow
---------------refinements
8
Product Line Architectures
blue-print for creating families of related applications
acknowledges companies don’t build individual products, but rather product families
importance: amortize costs of software design and development over multiple products
• most innovative work on software design next 10 years
• motivation not new: McIlroy ‘69, Parnas “families” ‘76• now! Jacobsson and Griss variation points
9
From Components to Refinements
Newest OO methodologies not based purely on objects/classes but on components
components are encapsulated suites of classes; scaling unit of design to packages/frameworks
ex: Catalysis, Rational advocating “component-based software designs”
OO, COM, Corba, Java Beans...
10
Perspective
I’ve been working on component-based PLA design methodologies for 15 years
• encountered domains where components cannot be implemented as OO packages, COM, Corba, …
• because performance would be horrendous
Doesn’t mean that components can’t exist for these domains
• rather, components must be implemented differently
• ex: metaprograms, rule sets of program transformation systems
11
Generalization...
Today’s notion of components is too implementation oriented or implementation-specific
Key idea: separate component abstraction from its implementation
• OO, COM, …, metaprograms, program transformation systems are implementation technologies
Abstraction that unifies the spectrum of component implementations is:Refinement
12
What are Refinements?
Changes to an application (when adding a feature) are not localized
Refinements are central to a general theory of Product-Line Architectures
• abstracts away “component” implementation details yielding common set of abstractions for all domains/PLAs
application
13
Benefit of Refinements...
Significant conceptual economy
one way in which to conceptualize PLAs
many ways in which refinements can be implemented
– choice is often domain-specific
conceptualbuilding-blocks
of PLAs
14
Towards Software Plug-and-Play...
Programming with today’s components is analogous to (old-fashioned) wire-wrapping hardware chips
traditional library paradigm
very successful, largely manual process
15
Plug-and-Play
Software construction allows for much greater degrees of automation
• want “intelligent” components to know how to “wire-wrap” themselves
• don’t manually specify myriad connections
Hardware Plug-and-Play
• standardized (hardware) interfaces• novices can do tasks of high-paid experts
16
Software Plug-and-Play
Do the same for software• standardized (software) interfaces• applications of PLA are specified/assembled as
compositions of components
• enable “average” programmers the ability to code like experts
Application#4 =
17
Motivations ...
Need for:
Product-Line Architectures Refinements Software Plug-and-Play
are clear...
18
Many Relevant Results...
extensible systems, open systems
domain-specific software architectures
aspect-oriented programming
subject-oriented programming
feature-oriented programming
generative programming
Note: approaches are NOT identical, essential problems they solve are similar
19
Common to Models of PLAs
define set of features that arise in a family of applications
each feature has 1+ implementations
an application specified by its set of features w. implementations
20
Classic Example of PLA
Booch Components 400+ data structure templates 18 varieties of dequeues = 3 x 3 x 2
Feature Implementation
concurrency (sequential, guarded, synchronized)
memory allocation (bounded, unbounded, dynamic)
ordering (unordered, ordered)
(and how not to implement PLAs)
21
Oops... Problems...
What happens when new feature added? ex: persistence
No conventional library could encompass enormous spectrum of data structures (or PLAs) that are encountered in practice
Library doubles in size!
22
Library Scalability Problem
n features ® product line of 2n applications
n features with m implementations ®
(1+m)n applications
Libraries of PLAs shouldn’t contain components that implement combinations of features
23
How to Implement PLAs
Libraries should contain components that implement individual, largely orthogonal features
component libraries are small O(n*m)
exponential numbers O((1+m)n) applications constructed from component compositions
ex: Singhal’s Components
24
Singhal Components
Reengineered Booch Components V1.47
Singhal Components• 1/4 size, larger product line• more efficient• easy to extend
Booch Singhal# of Components 82 22Lines of Code 11,067 2,760Domain Cardinality 169 208
25
P3 Generator
Generator of Java Container structures• 9 primitive data structures that can be composed
• ex: can generate a data structure where – elements stored in ascending order on field A,
and– hash-accessable on field B, and– key-accessable via red-black tree on field C…
• P3 equivalent to product-line or library of tens of thousands of containers
• generated software typically has faster execution times– optimize where static libraries cannot
26
What About Other Domains?
Lots of success storiesmostly independent
common set of ideas that are being reinvented
spend a few minutes explaining them...
Database Systems ProtocolsCompilers Radio SoftwareAvionics Audio Signal Processing
+ others
27
GenVoca
Simple, powerful, and abstract model of PLAs
• name derived from first PLAs based on this approach “Genesis” and “Avoca”
Takes idea of components that export and import “standardized” interfaces to its logical conclusion
28
GenVoca
Domain of applications = Product Line
has fundamental abstractions
define “standardized” interfaces (virtual machines) to abstractions
• may have multiple, interrelated classes• “virtual” because clients of interface don’t know
how interface is implemented
29
Realms
Set of components that implement the same virtual machine is realm
• is library of plug-compatible, interchangeable components
• lots of parameters - look only at realm parameters
S = { y, z, w }
R = { g[ x:S ], h[ x:S ], i[ x:S ] }
30
Parameters
Consider g[x:S] : R parameters define imported interfaces
g defines a refinement of R into S• refinement doesn’t depend on specific
implementation of S
interface R
interface Sg
31
Type Equations
Application is a named composition of components called a type equationS = { y, z, w }R = { g[ x:S ], h[ x:S ], i[ x:S ] }
A1 = g[ y ];
A2 = g[ w ];
A3 = h[ w ];
32
Grammars and Product Lines
Realms define grammars whose sentences are applications
Set of all sentences = product line
S = { y, z, w }R = { g[ x:S ], h[ x:S ], i[ x:S ] }
S := y | z | wR := g S | h S | i S
33
Recursion and Symmetry
Symmetric componentsexport & import same interfacecomposable in virtually arbitrary ordersof realm W have parameters of type W
ex: m[n[p]], n[m[p]], m[m[p]] ...
W = { m[ x:W ], n[ x:W ], p }
W := m W | n W | p
34
Why is Symmetry Important?
Applications can have open-ended sets of features
symmetric components are the way additional features are added to an application
not changing fundamental abstractions, only enriching them
35
Design Rules
Although equations may be type correct, there are always combinations of components that don’t make sense
semantic correctness of compositions
domain-specific constraints called design rules that preclude illegal component compositions
36
Product-Line (Domain) Model
Is an attribute grammarrealms of components define grammardesign rules are semantic constraints per rule
Generatorconfiguration-tool or compiler that implements
PL model
Type equation ApplicationGenerator
37
Model Says Nothing About...
When to compose refinements?dynamic (run-time)static (compile-time)
How to implement refinements?Mixins, OO packages, COM, Corba, …metaprogramsprogram transformation systems
38
Examples...
PLAs
OO…
MetaProgramming
Transformation Syst
Static
databasesavionicscompilers
data structures
protocols
Dynamic
protocols
audio signalprocessing
??
radio softw
are
39
What does this mean?
class foo { int a; int b; …}
class bar { …}
...
Classes!
Conceptualize software designs in layered/component-based refinements
Refinements!
TypeEquation
A1 = A[ B[ C ] ];
...
Designers who wanted to create product-line architectures by assembling customized applications via plug-and-play...
40
What is Gained?
PLAs of complexity and diversity that can’t be built in any other way
• handful of applications ® tens of thousands of apps
• performance of synthesized applications comparable to (usually better than) expert-coded software
• improved productivity: x 4 or more• 8-fold reduction in errors reported
Possible to build tools that automatically optimize equations (software designs)
• so novices can design and code like experts
41
But....
Problems and limitations with every approach
lots of technical problems, no “show stoppers”
hardest problems are non-technical
• typical of technology transfer
42
Non-Technical Problems
Legacy code• companies have legacy code that they want to reuse
in product-line applications• willing to accept penalty of hacking source code• reasonable for domains with little variation
Corporate politics• demonstrating PLA capabilities necessary but not
sufficient• egos, pre-existing methods, insecure funding …
can obscure technical goals• adoption decisions made for non-technical reasons;
results are often confused for technical reasons
43
Non-Technical Problems...
Think in terms of “layered” refinements and/or standardized interfaces
• greatest strength may be greatest weakness• architects may not be open to new approaches
Catch 22• “we won’t use it until they use it”
Terminology Arms Race• ex#1: “software architecture”• ex#2: Rational Software
44
Non-Technical Problems...
Not ready for prime time
“That’s not possible!”
“My software is too complicated to be built that way!”
45
Technical Problems
Open problem: testingcan synthesize applications quicklystill have to test applicationsnot clear how to reduce tests to reduce
product release time
Incompatible World ViewsBoston Gas Station Storyex: how to express refinements in OO?
46
OO Realizations of Refinements
A small-scale refinement adds new data members, methods, and/or overrides existing methods to a class
class
subclass
subclassing relationship
47
Large Scale Refinement
Adds new data members and methods simultaneously to several classes
class class class
subclass subclass subclass
48
Relationship to GenVoca
GenVoca components are consistent refinements of multiple classes
application classes
49
Scalability
application classes
Jak = blue[ black[ orange[ yellow ]]];
corresponds to over 500 classes, 26K lines of code
50
Technical Problems
Can express these ideas as mixins• I.e. a class whose superclass is specified by a
parameter
Want clean implementation in Java• neither Java nor Pizza supports parameterized
inheritance• need extensible Java (to add features to implement
refinements)
Jakarta Tool Suite (JTS)• PLA for Java dialects• GenVoca “generator” by which domain-specific
dialects of Java are assembled from components
51
JTS
(optional) features added to JavaLisp backquote/comma (to specify and
manipulate code fragments)hygienic macrosparameterized inheritanceP3 generator of container data structures…bootstrapped!
Free!Visit web site in paper...
52
Concluding Remarks
Heliocentricity was advanced in 1543, yet 60 years later it made no impact
people didn’t care about retrograde motions
Jean Bodin “No one in his senses or imbued with the slightest knowledge of physics will ever think that the earth staggers up and down around its own center and that of the sun… For if the earth were to be moved, we would see cities, fortresses, mountains thrown down… Arrows shot straight up or stones dropped from towers would not fall perpendicularly, but either ahead or behind…”
53
Concluding Remarks
How did heliocentricity take hold?telescope invented in early 1600s
• consistent with telescopic observations
provided simple explanation consistent with other disparate results• ex: earth tides
able to solve problems that were difficult or impossible otherwise
54
Concluding Remarks
This presentation motivated directions for future software development
product-line architectures
refinements as generalizations of components
codeless programming of software plug-and-play
GenVoca is one approach that has achieved all 3...
55
Concluding Remarks
When & how will “GenVoca” ideas take hold?
ideas constantly reinvented– plug-compatible components isn’t rocket science;– refinements aren’t new
lots of experimental evidence of power & capabilities
– much more in the future
reducing software complexity– standardizing abstractions of a domain/PLA is a
powerful way of managing and controlling the complexity of software in a family of applications
56
57
Timing
4:40 up to PLA 12:20 past how to achieve *14
min 20:20 up to GenVoca (20 min)
*22.5 min 29:30 up to interesting (10 min gen) 34:00 up to problems (17 finish)
*34 min
total: 47 min
top related