Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!rice!uw-beaver!Teknowledge.COM!unix!hplabs!hp-sdd!ncr-sd!ncrcae!hubcap!Donald.Lindsay
From: Donald.Lindsay@MATHOM.GANDALF.CS.CMU.EDU
Newsgroups: comp.parallel
Subject: Subdivision (was: scalability of n-cubes, meshes)
Message-ID: <7254@hubcap.clemson.edu>
Date: 30 Nov 89 13:38:02 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 26
Approved: parallel@hubcap.clemson.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.) 

Another design issue is the placement of the IO. A hypercube has 
n = log2(N) channels per node, so, you implement n+1 channels.  (For
n > 10, n and n+1 are indistinguishable.) Each node now has a direct
connection to an IO system. For example, NCUBE dedicates one 16-node
IO cube to each 128 nodes of the main cube. With a mesh, there are
fewer channels per node, with more wires each. Adding one more is
relatively more painful.  But if one doesn't, and (for example)
places the IO down the left-hand side, then the ability to subdivide
may be compromised.

Don		D.C.Lindsay 	Carnegie Mellon Computer Science


Brought to you by Super Global Mega Corp .com