Changing the Face of High Performance Code

1
Total Code Makeover

• Changing the Face of High Performance Code
• Brian Foote
• Wed., 18 May 2005
• University of Illinois at Urbana-Champaign
• http//www.laputan.org
• foote_at_cs.uiuc.edu

2
A Roadmap

• My Background, and Ralph Johnsons UIUC Software
  Architecture Group
• Objects, Frameworks, and HPC
• Refactoring HPC Applications with Photran and
  Eclipse

3
An HPC Carpetbagger?

• An Old Fortran Hand
• Realtime Programming
• Object-Oriented Frameworks
• Reuse
• Reflection
• Software Evolution
• Patterns
• Software Architecture
• High Performance Computing http//ibeam.cs.uiuc.ed
  u

4
The Gang of Four

• Design Patterns Elements of Reusable
  Object-Oriented Software
• Erich Gamma, Richard Helm, Ralph Johnson, and
  John Vlissides
• A landmark book that changed the way programmers
  think about building object-oriented programs

5
Patterns are Dispatches from the Trenches about
What Works

• Academics Feed the Novelty Vampires
• Secrets are Important to Industry

6
A Scene from a Nightmare
7
Big Ball of Mud

• The de-facto standard software architecture
• Why is the gap between what we preach and what we
  practice so large?

8
Sisyphus Revisited

• Writing variations of the same code becomes
  tedious
• Libraries work well, but have their limitations
• ObjectOriented Frameworks fill a void between
  libraries and applications

9
Frameworks

• Abstract Classes and Components that together
  constitute a generic or template solution to
  problems in a particular domain
• Do nothing well
• Often exhibit an inversion of control dont call
  me, Ill call you
• Extend via inheritance add fresh components
• Well know in GUI World others
• Can take a long time to learn should take a long
  time to build

10
Object-Oriented Frameworks

• White-Box Frameworks use Inheritance
• Black-Box Frameworks emphasize pluggable
  Components
• A skeleton, or driver, or main loop calls your
  application-specific code
• Polymorphism becomes the focus, not inheritance

11
Designing to Facilitate Change

• Code evolves in an iterative, incremental fashion
• Refactoring is the engine of the evolution
• Mature code exhibits more black box reuse
• Maturing code sports more patterns than immature
  code

12
Design of Scientific Applications(From Michael
Frumkinhttp//wiki.cs.uiuc.edu/PatternsOfHPC)

• Time represented as an outer loop
• Iterations over time step
• Space is represented by structured/unstructured
  grids
• Important for understanding data locality
• Data access patterns
• Spatial parallelism
• Physics is represented by an operator at each
  grid point
• Data flow
• Operator level of parallelism/dependence

13
The Last to be Shod

• Frameworks have been slow in coming to HPC
• High Performance remains a craft a cottage
  industry
• Reuse and Performance often at odds
• The Chasm HPC People have been O-O Late Adopters

14
HPC Programming Languagesand Objects

• FORTRAN (77, Carter era, James VI)
• Fortran 90 (Bush the Elder, George II)
• Fortran 2003 (Bush the Younger, George III)
• ANSI C
• C
• Matlab and Friends
• Assembler?

15
The Imperial Vision

• Youve seen one Grid Mesh, youve seen em all
• Too hard for Physicists
• Hotspots are in the Inner Loops, defy
  conventional SW abstraction approaches

16
Interoperatbility and Monoculture

• Make them write it all in your favorite language
• Interoperability is easy to achieve
• Fortran / C / C
• You can build frameworks in C
• You can fake em in C, with some effort
• Fortran awaits F2K3 (.foo, .f00)

17
Hand Crafting

• Skilled Craftsmen
• Hand Tools
• Tedious and often Dangerous
• Best way to achieve superior quality
• We Live and Die in the Inner Loop

18
Macro Processing

• Use the C Preprocessor (cpp)
• Even Fortran indulges this
• Crude, but effective
• Not easy to scale

19
Generative Approaches

• Use a scripting language or preprocessor to
  generate code glue, or data
• Use in lieu of of Macros
• CCA, IBEAM

20
Languages and Patterns

• X10
• Chappel
• Reflection and Metalevel Architectures (aka
  Aspects?)
• What ever else Ive forgotten
• Patterns

21
HPC Frameworks Works in Progress

• Frameworks have succeeded on a modest scale in
  HPC
• Tension between performance and flexibility
  remains high
• Message passing and sharing concerns still
  clutter the code

22
Refactoring Tools

• First developed at UIUC for Smalltalk
• Made popular by Fowler, Kerievsky
• Coming soon to a HPC near you

23
Refactoring

• Behavior Preserving code transformation
• Rename
• Extract method
• Verifies preconditions
• Reconciles all occurrences
• More reliable than cut n paste, or search and
  replace

24
Photran A Fortran Plugin For Eclipse
25
Photran History

• Opdykes C Tool
• Refactory and the Refactoring Browser
• IBM Eclipse Innovation Grant
• IBM PERCS High Performance Computing Initiative
• NASA IBEAM Project
• CRefactory

26
Status November 2004

• Produced two MS Theses
• Brought aboard two PhD Candidates
• Repaired numerous parser bugs
• Greatly Enhanced Stability
• Implemented Outline Views
• Make / Run / Debug under Linux
• Polished the GUI

27
Status May 2004

• Hyperion Release (14 January 2005)
• Parsed Full IBEAM Codebase
• Support for g77, g95, debugging under Windows /
  Linux
• Support for Lahey Fortran Windows
• Eos (Eclipse 3.0.1) (4 February 2005)

28
Why Refactor Fortran

• Make code easier to understand
• Make code easier to reuse
• Improve structure
• Banish duplication
• Make code your own
• Incrementally Cultivate the Code (XP)
• Refactor to Components
• Refactor to Patterns

29
What would you pay for a tools like this?

• http//www.photran.org

30
Something Wonderful

• Tension between performance and flexibility is
  high
• Moore and Amdahl are conspiring to help
• Computational Science ? Science
• HP Computing ? Computing
• HPC Pioneers will lead the way
• Pattern mining needs to begin now

31
Grid-Cogeneration

• Cycles too cheap to meter?
• You are insatiable. Ill sell you my Idle
  Process, at the going rate
• A Billion is an interesting