Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84; site watrose.UUCP
Path: utzoo!watmath!watrose!cdshaw
From: cdshaw@watrose.UUCP (Chris Shaw)
Newsgroups: net.arch
Subject: Re: Cube designs vs. x,y,z bus
Message-ID: <7306@watrose.UUCP>
Date: Sun, 24-Feb-85 06:00:56 EST
Article-I.D.: watrose.7306
Posted: Sun Feb 24 06:00:56 1985
Date-Received: Wed, 27-Feb-85 03:05:41 EST
References: <48@pbear.UUCP> <268@oliveb.UUCP>
Reply-To: cdshaw@watrose.UUCP (Chris Shaw)
Organization: U of Waterloo, Ontario
Lines: 39


The whole point of point-to-point communication channels is to eliminate all
forms of bus contention that may occur between processors. Hence CALTECH's
use of a large, complicated backplane setup; and Intel's use of point-to-
point ethernet channels in its cube products.

The point-to-point setup allows data to be passed from one processor to the
next in a ring, like a systolic loop, or through any number of other patterns
that may be available given an n-dimensional hypercube. Ultimately, perhaps,
one may want point-to-point between all processors. This would require
N*(N-1)/2  channels/ethernets/whatever, given N processors.
This would allow any processor to send data direct to any other processor in the
machine. For 64 processors, this would require 2016 inter-processor wires,
with 63 I/O devices per processor board. This is excessive.

The hypercube allows n* 2 **(n-1)  wires, where there are 2**n processors.
For 64 nodes, this amounts to 192 wires, with 6 I/O devices per board.
This arrangement can be built fairly easily (once the bugs are ironed out),
but you pay for it with access to only 6 of the 64 processors. However, access
to the remainder of the processors is a maximum of 6 communication links away.
The average communication length is 3, with no bus contention along the way.
The contention is really hidden in the point-to-point waiting for receive and
transmit.

The xyz bus doesn't deal with more than 8 processors very nicely, since you
still have bus contention possibilities on each of the busses, AND a wait
involved with changing from the x to y busses to get to a processor not
on the x bus. (x & y by way of example).

The x.y.z (or u.v.w.x.y.z) bus schema don't buy you anything. You get
buckets of bus squabbling for each "plane" AND you must route data to a 
different plane if the destination processor isn't connected to your plane.
Point-to-point, on the other hand, buys zero bus squabbling (i.e. full bus
bandwidth on each wire), at the price of data for "distant" processors
having to be routed through some intermediaries.

Hope this answers the "why point-to-point" questions ....

Chris Shaw