Subject: post-doc


Performance BBoard Message #511

Subject: post-doc

*******************************************************************************
==================================================================

POST-DOCTORAL POSITION
Performance Tools and Parallelizing Compiler Integration

BACKGROUND:

Historically
most distributed memory parallel systems were programmed
in single program multiple data (SPMD) mode using an explicit message
passing style
and standard workstation compilers were used to generate
code.  New data parallel languages like Fortran D and High-Performance
Fortran (HPF) express parallelism by specifying parallel operations on
arrays that have been distributed across the memories of the system.
Compilers for data parallel languages then create code that reads and
writes the distributed arrays using compiler-synthesized message
passing.  Instrumenting the data parallel source code will not reveal
the causes of poor performance; they lie in both the application source
code and the compiler-synthesized code.  Conversely
instrumenting the
compiler-synthesized code provides accurate performance data but no
mechanism to relate that data to source code constructs.

To provide performance feedback meaningful to a programmer writing in
an abstract parallel language such as HPF and Fortran D
performance
analysis tools must exploit information about the translation from the
high-level source language to the low-level parallel code.  The goal of
this ARPA-funded effort is to make it possible for programmers to pose
performance questions in terms of the machine-independent programming
model used in HPF and Fortran D and use these questions to acquire the
performance data needed to produce performance analyses and
visualizations that answer these questions in ways that are easily
related to the user's program.

This is a joint effort with Rice University and will involve close
collaboration with Ken Kennedy's Fortran D group at the Center for
Research on Parallel Computation.  It will build on the Pablo
performance instrumentation and data analysis software developed at the
University of Illinois and the Rice Fortran D compiler.  Our goal is to
explore the technical issues involved in integrating compiler and
performance analysis tools and to demonstrate prototypes of that
integration.

QUALIFICATIONS:

Recent Ph.D. with background and interest in experimental performance
analysis
knowledge of parallelizing compilers
and experience in
X/Motif and C++ development.

If interested
contact:

Professor Daniel A. Reed             Email:    reed@cs.uiuc.edu
Department of Computer Science       Phone:    (217) 333-3807
University of Illinois               Fax:      (217) 333-3501 1304
West Springfield Avenue
Urbana
Illinois 61801
USA

******************************************************************************