Hardware Architectures and I/O (was: Re: Jargon file...) **FLAME!!**

[Robert E. Seastrom]

In article <PST.90Dec1131440 at ack.Stanford.EDU> pst at ir.Stanford.EDU (Paul Traina) writes:
>
>Back when there were REAL(tm) computers like 780, a lot of time and
>energy went into designing efficient I/O from the CPU bus to the
>electrons going to the disk or tty.  
>

Damn right, but even the 780 was a step down.  Get your KL-10
documentation set out and read about *them*.  Front-end PDP-11s that
did Tops-20's command completion.  Seperate I/O and memory buses.
8-ported (that's eight, son) memory that talked to the I/O front-end
machines for *real* DMA, not cycle stealing!

>Sure OS's and apps have gotten bloated, but when you put a chip like
>the MIPS R3000 on a machine barely more advanced than an IBM-AT you
>end up with a toy that can think fast but can't do anything.  I can't
>really blame companies like DEC and Sun for producing mismatched
>hardware, because their marketing drones are constantly trying to
>undercut each other in price.  It's a hell of a lot more expensive to
>ship a product with a well designed I/O system than to drop in a
>"killer bitchen" CPU chip; occasionally someone makes the attempt do
>design a great piece of hardware, and you end up with something not
>half bad (like the DECstation 5000, which is only crippled by Ultrix

You left out the worst offender of them all - IBM.  The RS-6000 may
crank out 27 MIPS, but it can't context switch or handle interrupts
worth shit.  You can lower machine performance to the point of
unusability by FTPing a file from another machine on the same ethernet
segment!  Next time get a chance to play with an RS-6000, try this:
Pop about a dozen xterms, iconify them, put the icons in a row, and
wave the pointer back and forth over them as fast as you can.
Astounding, no?  The highlighting on the icons will keep bouncing back
and forth long after you stop waving the pointer.  My personal record
is 20 seconds.  Makes a Sun-2 running display Postscript seem
astoundingly fast.  RS-6000s also have an annoying tendency to "lock
up" for a few seconds (5 < x < 15) and then return to normal - I'm
told that this is normal and due to paging activity.  The microchannel
card cage design is pretty bad too - sure, you can put cards in, but
God help you if you have to take them back out!  And you better
tighten down the retaining screws all the way...  or the first time
you look at the card funny it will pop out.  To its credit, I must say
it compiles GNU Emacs faster than any other machine I've used, but I
do more with a workstation than just run compiles.  And, if you think
Ultrix is bad, it's only because you haven't tried AIX.

                                        ---Rob


-- 
Internet: rs at eddie.mit.edu             |  Copyright:  Protecting your right to
Bitnet:   RS at SESTAK                    |              copy software.
X.25:     PSI%0240200101905::KICKI::RS |                   ---gumby at cygnus.com