Exploiting object locality in JavaParty, a distributed computing environment for workstation clusters
JavaParty is a distributed computing environment that provides
transparent remote objects on workstation clusters.
Transparency provides the ability to defer object distribution
decisions and to make automatic object distribution possible.
For an efficient program execution in a distributed address
space, fast remote accesses are absolutely critical.
With KaRMI we have already shown how to realize that in pure
Java. But in an environment with transparent remote objects
fast remote access is not sufficient for efficiency. A good
object distribution will try to group objects according to their
communication patterns, not only expecting that local
communication is much cheaper than remote access but in addition
that local communication has an overhead similar to a normal
Java method invocation. Using RMI this assumption is void. RMI
does not exploit locality at all, because it was designed for
client server communication in a wide area network. KaRMI does
already exploit locality by providing a local transport
technology that shortcuts invocations within the same address
space. In this paper we show that detecting locality at the
transport level is too late and we present techniques to close
the gap between a Java method invocation and a local JavaParty
one. We conclude the paper with a kernel method invocation
benchmark that verifies the results and we demonstrate that
transparent remote objects can be used to efficiently execute an
application benchmark designed for a symmetric multiprocessor
|Zugehörige Institution(en) am KIT
||Institut für Programmstrukturen und Datenorganisation (IPD)
KITopen ID: 30482001
||In: Proceedings of the 9th Workshop on Compilers for Parallel Computers, CPC 2001, Edinburgh, Great Britain 2001 [online].
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page