Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!usc!ucsd!ucsdhub!hp-sdd!ncr-sd!ncrcae!hubcap!wen-king
From: wen-king@csvax.caltech.edu (Wen-King Su)
Newsgroups: comp.parallel
Subject: Re:  Subdivision (was: scalability of n-cubes, meshes)
Message-ID: <7277@hubcap.clemson.edu>
Date: 1 Dec 89 13:55:04 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 46
Approved: parallel@hubcap.clemson.edu

>From: Donald.Lindsay@MATHOM.GANDALF.CS.CMU.EDU
<No one has mentioned what I consider to be the fatal flaw of toruses
>(tori?).  Specifically, they can't be timeshared ("spaceshared").
<
>Anyone who invests in a massively parallel machine would like to be
<able to run multiple small problems on it. Program development, for
>instance. After all, you wouldn't buy two: better to put all the
<money into a single machine. 
>
<A hypercube is nice because it can be treated as two independent
>subcubes, and recursively one can have many subcubes. The message
<traffic within each subcube has no effect on the other subcubes - a
>wonderful property. A mesh is similar: you can have independent
<rectangular regions. A torus just doesn't work this way. (Neither
>does a mesh of buses.) 

My first project involving space sharing started with the Cosmic
Environment (1984).  The experiment eventually lead Intel to provide
space sharing on their iPSC multicomputers.  The experiments at the
time were all conducted on cubes and space sharing was done by
partitioning a big cube into smaller subcubes.  What we have discovered
when wormhole routing was developed is that process placement has
became much less important.

For medium grain multicomputers, message latency between two adjacient
nodes is not measurably different form message latency between two that
are the furtherest apart.  Hence don't always have to preserve the
topology when forming sub-machines, and a sub-machine can be just a
random collection of nodes that satisfies the requirements of the
program. Placement of special nodes (IO nodes, FPA nodes, Disk nodes,
etc), is, likewise, much less important.  The dissociation of
programming from architecture helps to make multicomputer programming
just a little bit easier.  This concept of allocation is demonstrated
on the Symult 2010 multicomputers.

One thing that favors grids over cubes is that when a cube is
subdivided, the total throughput of each node is reduced because some
channels in each node becomes disused.  If we subdivide a grid into
smaller grids, the interior nodes retain full message throughput.  One
thing that makes grids look very good compared to torus is that if you
subdivide a torus you get grids.  So why not just use a grid to start
with.

/*------------------------------------------------------------------------*\
| Wen-King Su  wen-king@vlsi.caltech.edu  Caltech Corp of Cosmic Engineers |
\*------------------------------------------------------------------------*/


Brought to you by Super Global Mega Corp .com