Skip site navigation (1)Skip section navigation (2)
Date:      Tue, 28 Jul 1998 12:13:56 -0400 (EDT)
From:      Daniel Frasnelli <dfrasnel@csee.wvu.edu>
To:        dbader@eece.unm.edu (David A. Bader)
Cc:        ports@FreeBSD.ORG
Subject:   Re: Ports category submission (fwd)
Message-ID:  <199807281613.MAA24187@naur.cs.wvu.edu>
In-Reply-To: <199807281540.JAA18461@jalapeno.eece.unm.edu> from "David A. Bader" at "Jul 28, 98 09:40:43 am"

next in thread | previous in thread | raw e-mail | index | archive | help
> I would include DSM, DFS, etc., as parallel computing tools, but I
> certainly wouldn't call them a "true 'parallel cluster'". 
	Actually, I was making the point that taken individually they do not
constitute a parallel cluster (Yes, I agree with you), but a true parallel 
cluster features all of these facilities (DFS, DSM, etc.).  My intention was
more or less to present a project idea: Instead of implementing one or two 
features of clustering (message passing and/or shared memory) as (I believe)
the Beowulf folks do, create an integrated parallel environment that is 
layered on Net/FreeBSD. 

> Many current supercomputers are logically equivalent to a cluster of 
> workstations with high speed interconnect, where tasks in a program 
> communicate by passing messages (e.g. MPI). 
	Quite true.  That which supercomputers do with high speed 
interconnect, a NOW does with ethernet cards.  A layered environment would
allow you to implement something like a hypercube topology not in hardware
(as a CM5/SP3/etc.) but in a network-based NOW.  Seamless integration of
shared memory/node communication is easier to implement in hardware (for 
obvious reasons), but is more easily maintained with a high-level 
abstraction of hardware in software.  Adding another computational node
to the Origin2000 I work on is a greater hassle than plugging another
PC into a hub :-) 

> In my research for high performance computing, I develop the most 
> efficient algorithms which dictate use of message passing, rather than 
> virtual shared memory, for performance. 
	NP problems are mostly memory-bound (as data space and complexity 
increases) rather than bound by message passing efficiency, aren't they?
(Not to say that message passing is insignificant!)
> This category should be inclusive -- *ANY* tools (above
> and beyond standard OS and networking infrastructure) which enable
> multiple CPUs to cooperate together to solve a computational problem.
	I completely agree with you.  Sorry if I gave an impression other
than that! :-)

Best regards,
Daniel
---
Daniel J. Frasnelli             Imaging spectroscopy research
(dfrasnel@wvu.edu)              Remotely sensed data analyst
Ecologist                       Extreme backpacker


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-ports" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199807281613.MAA24187>