19 Jun
2018
19 Jun
'18
12:51 a.m.
On Mon, Jun 18, 2018 at 1:56 PM, Noel Chiappa <jnc(a)mercury.lcs.mit.edu>
wrote:
Apologies for the repeat, if I have set this to TUHS before (I know I have
mentioned it in other places). Somebody else on this list mentioned David
Brailsford YouTube Video. Which has more details, biut he had some of not
quite right. I had watched it an we were communicating. The actual story
behind the byte is a bit deeper than he describes in the lecture, which he
thanked me - as I have since introduced him to my old friend & colleague
Russ Robelen who was was the chief designer of the Model 50 and later lead
the ASC system with John Coche working for him. Brailsford is right about
results of byte addressing and I did think his lecture is excellent and you
can learn a great deal. That said, Russ tells the details of story like
this:
Gene Amdahl wanted the byte to be 6 bits and felt that 8 bits was
'wasteful' of his hardware. Amdahl also did not see why more than 24 bits
for a word was really needed and most computations used words of course.
4, 6-bit bytes in a word seemed satisfactory to Gene. Fred Brooks kept
kicking Amdahl out of his office and told him flatly - that until he came
back with things that were power's of 2, don't bother - we couldn't
program
it. The 32-bit word was a compromise, but note the original address was
only 24-bits (3, 8-bit bytes), although Brooks made sure all address
functions stored all 32-bits - which as Gordon Bell pointed out later was
the #1 thing that saved System 360 and made it last.
BTW:
Russ and
Ed Sussenguth invented speculative execution for the ACS system
a couple of years later
. I was
harassing him
last January,
because for 40 years we have been using his cool idea and it
came back to bite us at Intel. Here is the message cut/pasted from his
email for context:
"I see you are still leading a team at Intel
developing super computers and associated technologies. It certainly
is exciting times in high speed computing.
It brings back memories of my last work at IBM 50 years ago on the ACS
project. You know you are old when you publish in the IEEE Annals of
the History of Computing. One of the co-authors, Ed Sussenguth, passed
away before our paper was published
in
2016.
https://www.computer.og/csdl/mags/an/2016/01/man2016010060.html
<https://www.computer.org/csdl/mags/an/2016/01/man2016010060.html>
Some
of the work we did way back then has made the news in an unusual way
with the recent revelations on Spectre and Meltdown. I read the
‘Spectre Attacks: Exploiting Speculative Execution’ paper yesterday
trying to understand how speculative execution was being exploited.
At
ACS we were the first group at IBM to come up with the notion of the
Branch Table and other techniques for speeding up execution.
I wish you were closer. I’d do love to hear your views on the state of
computing today. I have a framed micrograph of the IBM Power 8 chip on
the wall in my office. In many ways the Power Series is an outgrowth
of ACS.
I still try to keep up with what is happening in my old field. The
recent advances by Google in Deep Learning are breathtaking to me.
Advances like AlphaGo Zero I never expected to see in my lifetime.
"
Just like I retold the Amdahl/Brooks story of the
8-bit byte and
Amdahl
thinking Brooks was nuts
Don't think I've heard that one?
But you can lose with that strategy too.
Multics had a lot of sub-systems re-written from the ground up over time,
and
the new ones were always better (faster, more efficient) - a common even
when
you have the experience/knowledge of the first pass.
Unfortunately, by that time it had the reputation as 'horribly slow and
inefficient', and in a lot of ways, never kicked that:
http://www.multicians.org/myths.html
Sigh, sometimes you can't win!
Yep - although I think that may have been a case of economics. Multics
for all its great ideas, was just a huge system, when Moores law started to
make smaller systems possible. So I think your comment about thinking
about what you need now and what you will need in the future was part of
the issue.
I look at Multics vs Unix the same way I look at TNC vs Beowulf clusters.
At the time, we did TNC, we worked really hard to nail the transparency
thing and we did. It was (is) awesome. But it cost. Tru64 (VMS and
other SSI) style clusters are not around today. The hack that is Beowulf
is what lived on. The key is that it was good enough and for most people,
that extra work we did to get rid of those seams just was not worth it.
And in the because Beowulf was economically successful, things were
implemented for it, that were never even considered for Tru64 and the SSI
style systems. To me, Multics and Unix have the same history.
Multics was (is) cool; but Unix is similar; but different and too the
lead. The key is that it was not a straight path. Once Unix took over,
history went in a different direction.
Clem