Subject: Re: NetBSD with 4mb ram
To: None <port-vax@netbsd.org>
From: Paul Vixie <paul@vix.com>
List: port-vax
Date: 05/25/2002 21:37:47
a warning up front -- i've had enough to drink tonight that this discussion
has begun to seem entertaining.  just hit delete.  please read no further.

> an idiot (that's you, all you millions and millions of people who are only 
> pretending to actually do real work...which of course can't happen without 
> the perfection displayed in <insert your favorite dead and buried CPU 
> architecture>) don't realize this". 

at first i really despised the pmax (decstation 3100) because it used a MIPS
R2000 and i found its assembly language impenetrable.  (i still do, except
the alpha is even moreso.)  "i'll just stick to the vax, so that when the C
compiler generates the wrong thing i'll be able to figure it out."  but then
i noticed that a DECstation 3100 cost $5,000 and ran kernel builds twice as
fast as my 8650 (which cost $350,000) and that there were more than 200 of
the little pmaxes on my network but only two of the 8650's.  so, i switched.

don't get me wrong.  all modern processors have ugly assembly languages and
that hasn't stopped mattering, because C compilers are still haphazard.

but here i sit typing this to you on a dual-pIII freebsd system with a very
ugly assembly language that i hope i never have to look at again (especially
since it didn't help me much to look at it the last couple of times.)

the thing that made the older 360/370's, PDP11's and VAXes, even DG/MV and
HP3000 _seem_ faster than the PC's of their day (in spite of the fact that
my Z80B had a faster clock rate) was that they could do several kinds of I/O
at the same time.  little computers usually have one I/O backplane and one
path into and out of main memory, and no matter how many gigamumbles per
second it can carry, it will (statistically speaking) almost always be
carrying something for somebody else at the precise nanosecond you want it
to carry _your_ data.  so, a single-PCI machine will never be as good at
running hundreds of jobs "simultaneously" as an array-IO machine.  that's
why sun makes an enterprise 20000, and why DEC (compaq? hp? whatever) makes
an alpha ES40, and so on.  machines that can multiflow enough I/O for a
hundred "simultaneous" jobs will never be cheap enough to own (or quiet
enough to put under your desk).

but my favorite dead and buried CPU architecture is still a toss-up between
the fairchild clipper, the NEC V60/V70, and the intel 432.