Presenters:
Dr. Wolfgang Karl
http://wwwbode.in.tum.de/~karlw/
email: karlw@in.tum.de
Lehrstuhl für Rechnertechnik und Rechnerorganisation (LRR)
Institut für Informatik
Technische Universität München
D-80290 München
Germany
Tel: +49 89 289 28278
Fax: +49 89 289 28232
Martin Schulz
http://wwwbode.in.tum.de/~schulzm/
email: schulzm@in.tum.de
Lehrstuhl für Rechnertechnik und Rechnerorganisation (LRR)
Institut für Informatik
Technische Universität München
D-80290 München
Germany
Tel: +49 89 289 28399
Fax: +49 89 289 28232
For last few years the
SMiLE
(Shared Memory in a LAN-like Environment) project at
LRR-TUM
has investigated in the various aspects of SCI based clusters focusing on both hardware and
software developments. As one result of these efforts, a
comprehensive
software infrastructure has been created and compiled supporting
both message passing and shared memory style programming.
The latter one is based directly on the hardware DSM capabilities offered by SCI in order to fully exploit the performance potential of
the network and is offered to the user at two different levels of
abstraction: based on individually mapped segments of remote memory or
using a transparent global virtual memory system. While the first is
based on a straightforward implementation, the SISCI low-level API,
but burdens the programmer with extra complexity, the second one
provides a shared memory environment at a similar abstraction than
SMPs, but requires a complex hybrid DSM system. Such a system, the
SCI Virtual Memory or SCI-VM has been developed within SMiLE. Together with a general
framework for arbitrary shared memory programming models, called
HAMSTER (Hybrid-dsm based Modular and
Adaptive Shared memory archiTEctuRe), it provides the means for an
efficient and easy-to-use shared memory programming environment.
The presentation will introduce both models and describe their implementation, challenges, and applicability for the end user. It
will discuss how they relate to each other and how they fit into the
overall software infrastructure. This discussion will be based on
many examples and will enable the participants to understand the
issues involved in using SCI as the basis for shared memory programming.