Newsgroups: comp.os.parix
From: mbarnett@kingi.cs.uidaho.edu (Michael Barnett)
Subject: Topology Question
Organization: Laboratory for Applied Logic
Date: 06 Nov 1994 20:33:20 GMT
Message-ID: <MBARNETT.94Nov6123320@kingi.cs.uidaho.edu>

My interest is in creating library routines for global communication
algorithms. I have performed some work on machines like the Intel
Delta (and Paragon). Now I am interested in seeing if our techniques
work for machines like the Parsytec. But I am baffled by the
differences between the programming models. The Intel machines provide
a single method of message passing: the machine presents the illusion
that all nodes are directly connected, and one sends a message from
any node to any other node. The software overhead of sending messages
is high enough that one can ignore the difference between sending to a
physical neighbor and sending across the machine to a distant node.

But the Parix OS seems to provide that as just one option (SendNode
and RecvNode). It has the longest latencies; a much faster method of
communication is provided by first setting up a user-defined topology
and then using different communication primitives that take advantage
of the topology (SendLink and RecvLink).

My question is then: how do programmers take advantage of library
routines?  For example, my library routine for doing a global
broadcast might use a minimum spanning tree, but the user that calls
the routine has created a ring topology for their program. Is there
any alternative to having the library routine spend the time to create
the topology which it requires?  Is there any possibility of querying
the operating system to find out which topologies the user may have
already created, and if so, to use that topology without impacting the
user once the routine is finished?

I will be glad to summarize any responses. Thank you in advance,

Michael Barnett
Assistant Professor
-- 
Laboratory for Applied Logic		tel: 208-885-5524
Department of Computer Science		fax: 208-885-6645
University of Idaho			email: mbarnett@cs.uidaho.edu
Moscow, Idaho 83844-1010		www: http://www.cs.uidaho.edu/

