Architecture Recovery (a special, and especially important type of architectural analysis)

Temporal Aspect of Architecture Design decisions are made and unmade over a system’s lifetime – At time t a system has only one architecture Prescriptive architecture (PA) captures design decisions made prior to system construction – as-designed Descriptive architecture (DA) describes how the system has been built – as-implemented

iRODS – Descriptive Architecture How Many Systems Start off iRODS – Prescriptive Architecture …and End up

What Happened? Software decay – Drift – introduction of design decisions into a system that are not encompassed or implied by its architectural design – Erosion – introduction of design decisions into a system that violate its architectural design Decaying systems begin to “smell”  More on this later…

Another Example Hadoop Distributed File System – Prescriptive Architecture HDFS – Descriptive Architecture

Hadoop – HDFS + MapReduce

Hadoop – Complete Architecture

Hadoop – Complete Architecture, Another View

Architectural Decay Can exist both in the design and code Software smell Commonly made design or implementation decision Negatively impacts your system’s lifecycle properties It is not a bug – it doesn’t break your system It is a manifestation of technical debt Inspired by Refactoring: Improving the Design of Existing Code by Martin Fowler

A Catalogue of Architectural Smells Brick Concern Overload Brick Use Overload Brick Dependency Cycle Unused Interface Ambiguous Interface Duplicate Component Functionality Scattered Functionality Component Envy Connector Envy Connector Chain Extraneous Adjacent Connector …

Examples of Smells from Real Systems This document contains no technical data subject to the EAR or the ITAR.

Linux Architecture

Linux – Memory Manager Subsystem

Bash Architecture

Bash – Job Control Component

Bash – Commands Component

Hadoop – Complete Architecture

Hadoop – Dependency Cycle

Hadoop – Component Use Overload

Hadoop – Brick Concern Overload Value Aggregator

Hadoop – * Envy InterDataNode Protocol

What Can Be Done? Architectural recovery and then refactoring – The process of determining a system’s architecture from its implementation-level artifacts – Source code, executable files, Java.class files, … Difficult in practice – Size of code bases – Irrelevant details – Misleading details – Missing information – Hard to objectively assess existing techniques Still, many automated solutions available

Example: ShareDraw – Recovered Application Class Diagram

Clustering ShareDraw’s Implementation – Isolated Classes

Clustering ShareDraw’s Implementation – Generalization, Aggregation, Composition

Clustering ShareDraw’s Implementation – Two-Way Associations

Clustering ShareDraw’s Implementation – Recovered Partial Configuration FrameWindows_Mgr Dialog_Mgr View_Mgr Summary_Mgr DrawDocument_Mgr +m_pSummInfo DrawTool_Mgr DrawObj_Mgr Server

Existing Recovery Techniques ACDC – Algorithm for Comprehension-Driven Clustering – Structural pattern-based clustering ARC – Architecture Recovery Using Concerns – Concern-based hierarchical clustering based on similarity measure Bunch-NAHC & Bunch-SAHC – Hill-climbing algorithm for maximizing Modularization Quality LIMBO – scaLable InforMation BOttleneck – Probabilistic hierarchical clustering WCA-UE & WCA-UENM – Weigted Combined Algorithm – Dependency-based hierarchical clustering ZBR – Zone-Based Recovery – Hierarchical clustering based on textual information

Multiple Views Bash from ACDC Bash from BunchBash from ZBR Bash “Ground Truth”

Empirical Comparison Eight architectures of six open-source systems Obtained ground-truths for each ArchStudio 4 IDE Java 280K 54 comp. Bash OS Shell C 70K 25 Hadoop Data Prc Java 200K 68 Linux-C OS C 750K 7 Linux-D OS C 750K 120 Mozilla-C 1.3 Browser C/C++ 4M 10 Mozilla-D 1.3 Browser C/C++ 4M 233 OODT 0.2 Data Mgt Java 180K 217

Proximity to Ground Truth

Average Proximity 45%

Cluster Comparison

Average Cluster Similarity 20%

What Does This Mean? Architecture is multi-faceted Multiple ground-truths per system Different abstraction levels may be appropriate Architects may not really know what goes on in their system