Flash mob computing

From: Patrick Finnegan <pat_at_computer-refuge.org>
Date: Mon Feb 23 16:40:24 2004

Vintage Computer Festival declared on Monday 23 February 2004 05:03 pm:
> On Mon, 23 Feb 2004, William Donzelli wrote:
> > Supercomputers are dead. Or at least the magic aura of
> > supercomputers.
>
> Sure, but massively parallel is still quite vivacious as far as I can
> tell.

Supercomputers are following a trend we've seen already with "office
computing". Office-computing as I'll call it for lack of a better
name, started out as the IBM mainframe (or Unisys or Univac or ...)
sitting in the machine room at the 'center' of the corporate
infrastructure, and was gradually more attached to users; still the
vast majority of number crunching stayed in one central location. Next
was the outflux of independent (non-interconnected or loosly-connected)
PCs that at most had some ethernet connecting them together, in a
mostly peer-to-peer type fashion, with the situation now at a point
where there's vast tracts of PCs with some high-horsepower 'magic glue'
at the center holding them together (stuff like print servers, disk
servers, web servers, etc).

Likewise, we've seen Supercomputing go from very centralized machines -
vector processors like the Cray-1, Astronautics ZS-1, CDC Cyber 205,
etc. - to clusters of machines such as the ever-present (typically
Linux-based) clusters of machines that are fairly symmetric and don't
have a whole lot of glue, other than shared filesystems and network
switches, between them. This stage is like the outflux of PCs in the
80s to early 90s. This also includes other things like IBM's SP-type
systems, Sequent (now IBM) NUMA-Q stuff, the Sun F6800's we've now got
at work with their "Fire Link" intereconnect, etc.

The next stage (which some are working on, like the people at Cray) is
going to be a more asymmetric distributed machine, looking somewhat
like the current corporate IT infrastructure of a bunch of remote PCs
or thin clients that are fairly dumb with some high-horsepower
backends[1]. Basically, you'll have lots of small, fairly stupid
processors that are tightly coupled to some segment of memory, with
some smaller number of "master" processors that "farm out" jobs to them
to do on their segment of memory. So, basically, a tree-looking
infrastructure with memory at the bottom, memory coupled processors
above that, and above that system controllers that control things like
I/O and manage the memory-coupled processors.

So far, IBM is coming somewhat close to this model with their BlueGene
systems that they're working on, and Cray has a project that is funded
by DARPA (or someone else at that level of the government) which is
going in that direction - Burton Smith (if I've got his name
correctly), the senior engineer from Cray, came to Purdue to discuss
this project towards the end of last year.

Well, enough looking into the crystal ball for now. Time for me to get
back to studying/planning what to do with my 11/750.

Pat
-- 
Purdue University ITAP/RCS        --- http://www.itap.purdue.edu/rcs/
The Computer Refuge               --- http://computer-refuge.org
Received on Mon Feb 23 2004 - 16:40:24 GMT

This archive was generated by hypermail 2.3.0 : Fri Oct 10 2014 - 23:36:43 BST