=ADD=
=reftype=
14
=number=
01-12
=url=
ftp://ftp.risc.uni-linz.ac.at/pub/techreports/2001/01-12.ps.gz
=year=
2001
=month=
04
=author=
Schreiner; Wolfgang
=title=
Manager-Worker Parallelism versus Dataflow in a Distributed Computer Algebra System
=abstract=
We analyze two implementation variants of a parallel computer algebra
algorithm in Distributed Maple. The original solution uses a manager-worker
mechanism to control task scheduling, which requires an elaborate
administration scheme. The new algorithm is based on a dataflow approach
where all tasks are immediately started, automatically scheduled by the
runtime system, and implicitly synchronized by task dependencies;
non-determinism is effectively applied to provide more potential for
parallelism. It turns out that the new version is not only more declarative
(closer to the mathematical problem description) but also more efficient
than the original solution.
=howpublished=
PaCT'2001,Parallel Computing Technologies, Sixth International Conference, September 3--7, 2001, Novosibirsk, Russia. Lecture Notes in Computer Science, Springer, Berlin, 15 pages.
=sponsor=
Supported by grant SFB F013/1304 of the Austrian Science Foundation (FWF)
=keywords=
parallel computing, cluster computing, Distributed Maple, CASA, Java algebraic geometry.