Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!clyde!uunet!ncrlnk!ncrcae!hubcap!raja
From: raja@twilight.osc.edu (Raja Daoud)
Newsgroups: comp.parallel
Subject: Re: Subdivision (was: scalability of n-cubes, meshes)
Message-ID: <7323@hubcap.clemson.edu>
Date: 5 Dec 89 00:40:09 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 33
Approved: parallel@hubcap.clemson.edu

In article <7299@hubcap.clemson.edu> js7a+@andrew.cmu.edu (James Price Salsman) writes:
>
>I think the best solution for MIMD configurations is
>dynamic reconfigurability, such as provided by a big
>software controlled switch.  The INMOS C004 does this
>for transputer networks, without much loss of link
>bandwidth.
>
A practical rule of thumb (i.e. measured on real hardware):
A message passing through 3 C004 chips looses 1/2 bandwidth
due to bit re-synchronization within each C004.  That limits
(on the practical side) the scalability of a "Big Switch" made
out of many C004 chips.  There still is the question of "who"
controls the switch (centralized/decentralized) and how many 
CPU cycles are wasted making the reconfiguration (decision and
letting each CPU know where it stands (routing ...)) because 
users of MIMD machines want "application speed" at the end of
the day, and if dynamic reconfiguration takes more time than
what is gained by doing it, it wouldn't be practical (Not with
the current hardware at least).
On the other hand, software solutions that require little or no
topology information are IMHO more attractive in the long run
and easier for hardware to support.

>:James "I'm *SO* sick of 3D" Salsman

--Raja

-=-
-----------------------------------------------------------------
Raja Daoud				raja@tbag.osc.edu
Trollius Operating System		(614) 292-4122
The Ohio State University		Ohio Supercomputer Center