Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!clyde.concordia.ca!uunet!lll-winken!brutus.cs.uiuc.edu!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!emory!hubcap!birenboi%girtab.usc.edu
From: birenboi%girtab.usc.edu@usc.edu (Aaron Birenboim)
Newsgroups: comp.parallel
Subject: Re: Subdivision (was: scalability of n-cubes, meshes)
Summary: What about FLIT routing?
Message-ID: <7432@hubcap.clemson.edu>
Date: 14 Dec 89 15:46:07 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 34
Approved: parallel@hubcap.clemson.edu

In article <7254@hubcap.clemson.edu> Donald.Lindsay@MATHOM.GANDALF.CS.CMU.EDU writes:
>No one has mentioned what I consider to be the fatal flaw of toruses
>(tori?).  Specifically, they can't be timeshared ("spaceshared").

Why not?  I believe the NCUBE sets up a kind of virtual sub-cube
topology.  You refer to a sun-cube's nodes by a "pointer" index.
the O.S. decodes this into the actual node location, so a sub cube of size 
2^r may not actually have nodes where the node address the programmer uses
and thinks of as a nearest neighbor is actually a nearest neighbor.

Why not have the tours' or grid's O.S. parse out "virtual sub-tori"
which are not physical tori, but look like it to programmers.
The O.S. might have some trouble with assigning nodes which comminicate
a lot close to each other, but this may not be a large problem.

Is the discussion on grids and tori considering flit routing
a la NCUBE?  By sending out a flit and establishing a communication
path all the way fron source node to receive node, ferthest neighbor
communication is only 10% longer than nearest neighbor (an NCUBE
claim).  I realize that routing the connection time depends
linearly on the number of nodes traversed, but the actual message
transmission does not.

Were people keeping the flit routing in mind when they made the
claum that a return to hypercubes would be a good decision for
optical interconnection networks?  If so, why?  The flit routing
will take a similar amount of time, the messagge transmission
is dependant on simple inter-node bandwidth, number of hops
do not have such an extreme effect.
--
=============================================================================
Aaron Birenboim   |  aaronb@eedsp.gatech.edu
GT Box 30735      |
Atlanta, GA 30332 |  (only temporarily at birenboi@girtab.usc.edu)