[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [f-cpu] mCLUSTER

On Fri, 7 Sep 2001, Richard E. Hartny wrote:

> Couple things for Juergen---
> You stated - How about a bottleneck analysis of the applications?
> We have.  The bottleneck was - Process Bound and I/O bound with the disk.
> Current implementation cures these ailments.  The ideal situation now is to become I/O bound - hard disk that is.
> I too worked with 32 TTY' s (64) to be exact (32 for hardware design and 32 for Software Design..  All into a processor front-end to to the Honeywell - number forgot. From these inputs we received about 4 hours later a Mag Tape we used for Floating Point firmware (micro code as now called) or Tape for hardware.  Happy - NO. Twiddle'd our thumbs waiting.
> Now for for what I think you are trying to say about a distributed system
> In 1976-78 I worked on a system called TASES (Tactical Airborne Surveilance Exploitation System)  It consisted of four Processors ( an Emulation of the 
> NOVA-3 - Data General Mini). There was an IF (Intermediate Frequency) processor, an Audio Processor, and Azimuth Processor and a DF (Direction Finding) Processor.  Prior to this system it a minute or two to get a DF.  This was reduced to approximately 50 milli-seconds. It was also determined FRIEND or FOE.  Russian or one of ours.  What made this all possible was a Programmer discription to me of an FFT Transform.  A lot of Multiplies and Add.  So I investigated MUL for a good six months.  Ended up with what I called an AFG Board.  AFG = arithmetic function generator.  Used a Time- sequence MUL, Bit Reversal, Square Root PROM look-up, a Trig function I forget and a Memory containing Antenna Correlation Coeficients.  The audio was recorded and the Pilot of the Aircraft was the lone human interface. With several Aircraft you pinpoint, up or down, and where is it.  Ours or theirs.  Missle away.....
> Am I on target - or is their more?????


quite interesting and I didn't mean to let you down before.
There are a lot of views one can take of systems. In a cpu
design it's only necessary to have a fast switching time to
serve external events without loosing the many things done
for the 'main' task (e.g. register, cache contence...). But
even this can be improved or equalled by a buffered/cached
IO subsystem design that even may use some local intelligence.

By the way mCluster is supposed to have up to 1000 cpu nodes
in a single qubic meter but it's not sure that it will ever
come with f-cpu. ;-)


To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/