Overview of the D System Project
Rationale
Supporting high-level, machine-independent parallel programming
requires:
- aggressive parallelizing compilers with advanced program analysis and
code generation capabilities; and
- tools that enable a programmer to develop and tune programs at the
source language level, without having to understand the complex,
machine-dependent code generated by a parallelizing compiler.
Achieving both these goals requires sophisticated compilation technology
as well as close cooperation between parallelizing compilers and tools
in order to support source-level application development.
Components
The Fortran D95 language
An experimental language based on High Performance Fortran,
with extensions under development to support important
classes of parallel applications not well supported under
current HPF:
- out-of-core computations on large data sets
- irregular data layouts for unstructured and sparse matrix
computations
- task parallelism for multi-disciplinary codes and
explicitly parallel algorithms.
A new, second-generation, data-parallel compiler that
is the foundation for our ongoing compiler research.
More information is available on the dHPF Home Page.
The D Editor *
An intelligent browser for Fortran D95 that provides
feedback on the analysis and parallelization performed by the
Fortran D compiler;
The dPablo Performance Tool *
A dynamic performance measurement and visualization tool integrated
with the D Editor, to provide source-level explanations of measured
program performance. This is joint work with the
Pablo project
at the University of Illinois.
Automatic data layout techniques *
We have developed techniques based on performance prediction
and 0-1 integer programming to choose optimal data layouts for
a data-parallel program. This research provides the foundation
for a tool that can assist a user in data layout selection.
FIAT: Framework for Interprocedural Analysis and Transformation
An interprocedural program analysis system and repository to support
efficient whole-program analysis.
Participation in
POEMS
The POEMS project aims to design and prototype an environment for
predicting end-to-end performance characteristics of large-scale
parallel applications. This project involves researchers at several
other institutions.
Our group at Rice will focus on compiler-based techniques for
constructing performance models of HPF programs at different levels
of abstraction, from a detailed specification language to abstract
analytical models.
* These tools were originally based on the Fortran D compiler (our
previous generation compiler), but are currently being ported to dHPF.
If you have questions or comments about the D System or this repository,
please contact dsystem-info@cs.rice.edu.
Last updated on 24 July 1997.
URL: http://www.cs.rice.edu/~dsystem/overview.html