BROADWAY: A SOFTWARE ARCHITECTURE FOR SCIENTIFIC COMPUTING

1
BROADWAY A SOFTWARE ARCHITECTURE FOR SCIENTIFIC
COMPUTING

• Samuel Z. Guyer and Calvin Lin
• The University of Texas

2
MOTIVATION

• Efficient runtime algorithms and efficient memory
  usage are critical
• Temps programmers to emphasize performance over
  clarity
• Often complicates long term maintenance and
  portability of scientific code.
• Practitioners of scientific computing are
  typically unwilling to try novel languages or use
  sophisticated design methodologies
• Deep knowledge of scientific domain is required

3
BROADWAY COMPILER SYSTEM ARCHITECTURE
Library
Annotations
Header flies
Source Code
Integrated and optimized code
Broadway Compiler
Application Source code
4
Performance problems caused by use of software
libraries

• Different clients have different needs
• Separation of concerns inhibits information
  flow across interfaces
• Worst case assumptions provide generality at the
  expense of performance
• Modular structure leads to poor resource
  management
• No performance portability
• Premature optimization complicates maintenance
• Interface bloat defeats modularity

5
LIBRARY LEVEL OPITIMIZATIONS

• Replace general purpose routines with customized
  routines that run much faster.
• Can make stronger assumptions about calling
  context, allowing for more efficient execution
• Avoid interface bloat
• Use of annotated language to describe domain
  specific information
• Optimization requires application programmer to
  have little knowledge of library
• Broadway compiler can specialize library routines
  in ways that the library designer did not foresee.

6
BROADWAY AS A META INTERFACE

• Use of black boxes leads to performance problems
• Implementation and interface that black boxes
  provide will inevitably be inappropriate for some
  client

Application
Base Interface
Black Box
7
BROADWAY AS A META INTERFACE

• Solution provide 2 interfaces, a base, which
  most clients use, and a separate meta interface,
  which allows the black box to adapt to the needs
  of different clients

Application
Base Interface
Black Box
Meta- Interface
Meta- Program
8
Experimental Results
Benefit of Automated Approach
Tedious transformations can be applied uniformly
and completely!
9
Conclusion

• Provided mechanism for improving performance
• Create different library implementations and
  interfaces for different clients
• Conveys library-specific information to the
  compiler and use this information to customize
  the library for different users.
• Replaces invocations of general routines with
  invocations of specialized routines, thereby
  relaxing worst case assumptions.
• Improves the maintenance and portability of
  applications
• Enhances the value of legacy codes
• Provides clean division of labor

10
Discussion

• What are the possible downfalls?
• Meta-Object Protocols
• Provide mechanism to change the way a language is
  compiled.
• Future optimizations?