Ph.D. Thesis: Scoped Behaviour for Optimized Distributed Data Sharing
Copyright 2000. All rights reserved.
introduce the novel scoped behaviour abstraction
and examine how it is used to optimize distributed
data-sharing patterns within
parallel programming system.
Scoped behaviour is an application programmer's interface
to a set of system-provided optimizations; it is also
an implementation framework for the optimizations.
is a distributed shared data system
where shared-data objects are implemented in C++
as abstract data types.
has been prototyped on a network of workstations
connected by an ATM network.
design, implementation, and evaluation of
scoped behaviour contributes to the field of parallel and distributed systems
by demonstrating that:
- Scoped behaviour can provide per-object and
per-context (i.e., specific portion of the source code) flexibility
when applying data-sharing optimizations.
In contrast to some other systems,
programs can be incrementally tuned
and only a minimum number
of error-prone changes to the source code are required in order to
experiment with different optimization strategies.
- Scoped behaviour can be implemented
without language extensions
and without special compiler support.
Scoped behaviour's novel implementation framework
can exploit both compile-time and run-time information
about the parallel program.
- A parallel programming system based on a high-level
shared-data abstraction can achieve high performance.
In a performance evaluation of four applications implemented
using three different types of parallel programming systems,
usually matches or
outperforms, sometimes by a wide margin,
TreadMarks (a distributed shared memory system) and
a message-passing system (either MPICH or PVM, depending on
- Abstract only.
Gzip'ed Postscript (17 kbytes)
- Complete thesis (single Postscript file, 203 pages).
Gzip'ed (300 kbytes)
Return to home page
$Id: thesis.html,v 1.3 2000/07/18 00:38:29 paullu Exp paullu $