[Acpc-l] Vortrag HEUTE : M. Malawski and K. Zajac
Peter Brezany
Peter Brezany <brezany@par.univie.ac.at>
Mon, 24 Sep 2001 14:58:59 +0200 (MEST)
UNIVERSITAET WIEN INSTITUT FUER SOFTWAREWISSENSCHAFT
gemeinsam mit
FWF-Projekt Spezialforschungsbereich F011 "AURORA"
EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES AURORA-KOLLOQUIUMS
ZEIT: Montag, 24. 9. 2001, 16.15 Uhr s.t.
ORT: Institut fuer Softwarewissenschaft,
1090 Wien, Liechtensteinstrasse 22,
Seminarraum, Mezzanin
Maciej Malawski and Katarzyna Zajac
Institute of Computer Science, AGH
Krakow, Poland
Advanced Library Support for Irregular Data Intensive
Scientific Computing on Clusters
In the past few years, clusters of workstations and PCs have emerged
as an important and rapidly expanding approach to parallel computing.
Although current cluster systems involve large main memories, many
advanced applications access data sets that are often too massive to
fit in main memory of even the most powerful clusters and must therefore
reside on disk and be fetched during execution of the program.
Such applications are called out-of-core (OOC) applications. Traditionally,
in scientific computations, OOC problems are handled in two different
ways: (1) virtual memory, which allows the in-core program version to be run
on larger data sets, and (2) specific OOC techniques explicitly interface
file I/O and focus on its optimization.
This talk presents an advanced library system called LIP supporting
parallelization of irregular codes, an important class of scientific
applications, for cluster systems. The library supports parallelization
of both out-of-core and in-core irregular applications in Java,
C, and C++ programming environments and is built up on top of
MPI/MPI-IO. Therefore, it is fully portable.
The talk is organized as follows: first, we introduce
the terms and concepts relevant to the area of irregular and OOC computing.
Next, we outline the parallelization and optimization techniques for
irregular problems. We present the classification of OOC problems
and describe optimization methods used in this field.
Then, the main concepts the LIP library is based on are presented.
We also outline a Java to Native Interface Extension
Tool that was used to create Java bindings to the LIP library.
We present performance results obtained from runs of synthetic
and real benchmarks on various parallel architectures.
Finally, we summarize the results of our research and outline several proposals
for future development - the main focus is devoted to possible extensions
of the LIP library toward Grid computing.