🚀 go-pugleaf

lamaster@pioneer.arc.nasa.gov (Hugh LaMaster) writes:

>In article <1993Jan26.215541.9957@adobe.com>, zstern@adobe.com (Zalman Stern) writes:
>|> In article <1k46ioINNijv@fido.asd.sgi.com> gints@prophet.esd.sgi.com (Gints  
>|> Comments like this indicate that the big iron boys are pretty desperate.  
>|> They are fighting the dynamics of the market. There are *relatively* few  
>|> problems that require "mainframe power." Over time mainframes must either  
>|> become cheaper or more powerful. More powerful is limited because there are  
>|> fewer and fewer problems that require that kind of power and it becomes  
>|> harder and harder to make the machine more powerful. Cheaper is a problem  
>|> because the margins go down and a lot of a mainframe's costs are not  
>|> hardware oriented. (I.e. service, paying for the space and other central  
>|> facilities required.)

[stuff deleted] 

>The reason I am posting this is that I think it is a big mistake
>to underestimate the importance of big iron systems.  IMHO, most large
>data centers will continue to use large systems to provide the cheapest
>deliverable "bulk" CPU cycles, to manage large and growing data
>requirements, and to provide the highest possible transaction rates.
>As well as large engineering and scientific simulations.
>What has already disappeared is the use of "mainframes" for relatively
>mundane word processing, etc.  But, the data management requirements
>are still there.  Today, big iron systems have memory bandwidths of
>1-200 GBytes/sec, raw IO bandwidths of 1000 MBytes/sec., I/O
>throughputs of around 100-200 MBytes/sec and up to 2000 integer-scalar 
>MIPS.  While this is maybe only ~200 486 systems, there is an enormous 
>difference in the capability between such systems.  A 200:1 difference in 
>power totally changes the capability of a system, as was pointed out 
>recently at Supercomputing '92.  I could go on, but, let's not confuse
>the financial troubles of a few big iron manufacturers with the 
>disappearance of their market.

That is true. I have run some simulations on CONVEX machines, workstations,
and PC's. I have an interesting simulation which would take about 3 months
on my PC, one month on the fastest workstation, and two weeks on a CONVEX.
(pretty well vectorizing code with big datasets, as you may quess. The PC
and workstation times are estimates from partial simulations.)

The PC time is completely unusab|e, even if I divide the problem for one
hundred PC's, I'd not get much more than communication traffic between those
PC's. The workstation time is almost usable, and the fortnight on CONVEX is
clearly practically usefull. Not every problem can be distributed for a
number of PC's because the dependencies would require so much communication
that most of the cycles would be wasted for it.

I'm sure that there will always be need for machines that are 6+ times
faster than the fastest PC. Well, 100+ is always better :-). 

----------------------------------------------------------------------
Jarmo J. Ahonen
Computing Centre, Lappeenranta University of Technology, P.O.Box 20,
SF-53851 Lappeenranta, Finland. email: Jarmo.Ahonen@lut.fi