On Wed, 13 Jan 1999, Robert M. Hyatt wrote:
> since we are on the subject. our three supersparcs that died had 32meg
> memory modules (4 @ 32). When they died and Sun came out, they charged
> us 3 x 400 dollrs to replace one 32 meg DIMM in each machine. Yes, you
> heard that right, 400 bucks for a 32 meg DIMM. I just paid 300 bucks for
> 128M 72 bit ECC dimms for my quad xeon. When he gave me the bill, I held
> my hands over my head. He said "what's that all about?" I told
> him "if you are going to rob me, I'm damned sure going to hold my
> hands over my head so you don't also shoot me. :)
>
> He was pissed.
>
> I was pissed.
>
> Had to keep a mop handy for days in fact.
>
> ANd then he charged me 1000 bucks for a 524mb scsi disk (yes, 1/2
> gig).
>
> And I held my hands up when he walked in.. :)
This sort of stuff is why I gave up Sparcs. A simple story may be
instructive that isn't totally inappropriate for linux-smp as it points
out some of the reasons that linux has succeeded as strongly as it has.
It also shows why I am absolutely certain that linux (and freebsd;-)
will "eat Microsoft alive" over the next few years. Here are my
shameful confessions:
"My name is Rob Brown and I - uhh -- used to like Suns" he said, his
unshaved face grizzled with greying hairs, and his watery blue eyes
somehow forlorn as they swept across the understanding faces of the
folks sitting in the ring of chairs that surrounded him...
Yes folks, I used to be a rabid Sun-Lover, mostly because some years ago
SunOS 4.x.x was the BEST (Unix) operating system distribution available
on any platform -- it was the direct heir of BSD Unix, and Sun's
scheduler was tuned to provide excellent interactive performance even on
a heavily loaded system (given, of course, a 3-15 MIP processor). They
were also not all that expensive compared to the general run of
workstations with a standard 40% University discount.
Sun then made several REALLY DUMB strategic mistakes. Each of these
mistakes created market circumstances that would ultimately cost Sun
what by now has probably grown to close to $100 Billion. They also,
incidentally, led to the creation and evolution of linux. It is worth
understanding this.
First, there was a brief shining moment back in the mid to late 80's
where Sun had a version of SunOS that would run on Intel CPUs -- my
first Sun box was actually a Sun 386i. True, the 386i was a "custom"
piece of overpriced hardware with a number of significant flaws (tell me
about 'em!). Still, they had WRITTEN the core assembler for the CPU,
they had PORTED a C variant of the SunOS sources, they had a compiler,
they had built a standard set of applications (and pretty much anything
written for SunOS would build without tweaking on a 386i) and they had
device drivers for at least their custom SCSI interface, ethernet, and a
frame buffer. They were in a great position to build what, a handful of
device drivers (IDE, VGA, ethernet for at least one card) and sell a
version of SunOS that would run on a "standard" 386 PC with 4 or more
megabytes of memory for $50. Sunview and all.
Had they done so, Sun would "own" the software/OS world now, just like
Microsoft does instead. I mean, who would have bought Windows 3.0 with
no compiler, no software, a need to run on top of DOS, no real
multitasking, no real networking, etc. when they could have bought
SunOS, a real Unix, with ALL of the above and a huge library of public
domain applications even BEFORE the stampede ensued to write
applications? Nobody. Not corporations, not consumers. Sun could have
been the "Turbo Pascal" of mass market operating systems with an
absolutely insurmountable lead over Microsoft (or anybody else) and
would have made far more money from selling software than they ever made
from selling hardware.
Linux, of course, would never have been born, or if Linus had gone ahead
and written linux even with a cheap, popular, and universal Unix for
PCs, it would have had a much harder time attracting all the programmers
and talent that now contribute to it (at the time I'm speaking of, they
were all writing PD software that was typically built to run under
SunOS).
Instead, Sun gave Microsoft YEARS to build a real multitasking operating
system and smoothly transition an immense mass of corporate software and
systems from one shit operating system to another while gradually
increasing both cost, profit, and functionality. Talk about dumb.
Second, Sun failed to differentiate between "kernel" and "distribution"
when they released Solaris. I won't get into an argument about whether
Solaris is technically superior to SunOS. Frankly, SunOS was quite
excellent UP, and SunOS SMP was very similar to linux 2.0.x (single
kernel lock but excellent userspace multiprocesser-ing). I think that
the tremendous success of linux 2.0.x even today clearly shows that a
refined kernel spinlock, while clearly important, is far less important
than people imagine from the point of view of producing a stable and
usable and satisfactory system for MANY different purposes. And
besides, I could care less if Sun improved their kernel or even changed
kernel paradigms, as long as it continued to properly function.
So, Sun decided to convert to SysV-type kernels to get "threads". This
was just great by me -- if it boots and runs smooth the paradigm doesn't
matter to users or administrators, only to programmers (and not even to
all programmers since you don't HAVE to use threads in procedural
single-threaded code, which is very, very, common). However, they ALSO
decided to change their entire bloody operating system DISTRIBUTION over
to SysV. Even to this day, I have a hard time grasping this. They had
some 30-40% of the workstation market, they "owned" the University and
were making inroads into the corporation, and all those expert systems
administrators and systems programmers were overnight told that they had
to learn different flags for ls, for ps, learn a different layout for
/etc and a different boot sequence -- basically, they were told that
they would spend the next six months relearning to do what they were
already very, very good at doing but differently.
Let's face it. Software is cheap. Hardware is cheap. Human time is
expensive. Take some tens of thousands of systems personnel, charge
them a thousand hours (half a year FTE) apiece to convert skills, to
rebuild software, to struggle with a new kernel (which, as it happened,
sucked compared to SunOS on SINGLE processor systems up through Solaris
2.5 -- I don't know if 2.6+ are finally usable because with linux around
it's "Frankly, Scarlett, I don't give a damn") and you have saddled a
market known to be strapped for cash and time and intolerant of bullshit
with (say) 10^5x25x10^3 = $2.5 BILLION in conversion expenses. Except
that this still neglects the cost of supporting all the USERS of these
systems as THEY learn that ps now has different flags, sorry, and that
oh, that application isn't ported and rebuilt yet, sorry, and... So we
could easily multiply this figure by a factor of 2-4 and probably still
be short on the real cost of the SunOS->Solaris forced transition.
To avoid our little piece of this immense cost, we (like many, many
others) held out as long as we could, running SunOS on our aging ELC's,
SS2's, and SS 10's (after Sun realized that we were NOT going to run
Solaris, they were arm-twisted into releasing SunOS for each new
platform that was originally released for "Solaris only" until the
Ultra). We all began to look for alternatives, and surprise! There was
linux, all ready to go and running on commodity PC's that were just
getting to be as powerful as Sun workstations anyway. Furthermore, some
DISTRIBUTIONS of linux were refreshingly SunOS/BSD-like in their
interface. A PC running Slackware could literally not be told from a
Sparcstation 2 running SunOS at the command level, most /etc files were
where they belonged (or close to it, it made some concessions to SysV
but kept the hand-editable flatfile paradigm going) and was sufficiently
close to a Sun at the compiler/library level that "most" of that vast
library of PD software (obviously including all of gnu:-) could be
rebuilt on a linux box with only a few tweaks, far less than what was
required for Solaris. Indeed, the distributions came with a hell of a
lot of what one used to build and maintain by hand prebuilt FOR you in
the distribution. It was BETTER and EASIER than SunOS to install and
manage!
Well, add up the numbers. Cheap PC's (see below) save money on
hardware. Cheap (free) OS distributions that preserve the interface one
has grown to know and love permit one to avoid the huge time investment
required to change over to something different. The only remaining
question was, did linux work well enough? Was it stable (SunOS was also
legendary in its stability, routinely accumulating months of uptime
before it was wise to therapeutically reboot a system if it hadn't
crashed, probably from a minor memory leak or something)? Well, as we
all know it was at least as stable as SunOS back in 1.x, and although
there have certainly been unstable revisions, it has never been a
problem to find a stable revision and stick with it. Even with
2.0.(x<30) when the SMP kernel had KNOWN deadlocks in it, we could
usually rack up months of uptime without getting nailed by one.
The third and last mistake Sun made they are still making, as Robert's
little anecdote (which motivated this whole sad story of my 12 point
plan to kick the Sun habit) clearly shows. They have utterly failed in
the hardware market to respond to the advent of Intel systems (or
perhaps better, COMMODITY systems in general) as serious competitors in
the workstation/server market. To this day an Ultra workstation costs a
truly absurd amount compared to what one can get from a local dealer
with an Intel or AMD or Cyrix or DEC Alpha CPU in it (note well that
Digital did NOT fail to respond and made a better/faster CPU anyway).
Sun has maintained an immense price/performance barrier that effectively
excludes (sane) University departments from using their hardware except
in speciality applications or because they are rich enough to afford it
and don't give a damn. Of course these same Universities are training
the next generation of systems programmers, the next generation of
systems administrators, and so on. Some poor Universities have even
turned to NT, for gosh sake, just to be able to buy cheap enough
hardware that they can cover folks' desks. Time was, all those desks
had Sun boxes on them or plain old PC's. One cannot even afford to
maintain a Sun box.
Let's see, Robert, you paid $1000 for a 500 MB drive and $1200 for 3 32
MB DIMMS to keep your three aged Supersparcs alive. This gives you a
budget of $2200. Hmmm, if I walked into almost any computer store (even
Best Buy) I could buy over the counter at least 2 Celeron based PC's
with a MINIMUM of 12 GB between them, a MINIMUM of 128 MB of SDRAM
between them, NICs, 17" color monitors and accelerated graphics, and
even have enough leftover to buy nice multimedia kits with big speakers.
If you shopped carefully and were willing to add an extra $500-800, you
could replace all three with nice clean year warranties. I don't know
how many CPU's your SuperSparcs have, but it is almost dead certain that
one Celeron at 300 MHz would eat them alive (possibly all the CPUs in
all 3 systems alive).
Sun just doesn't get it. They make it IMPOSSIBLE to use their hardware
if you are a sane and reasonable person, even if you are (like I once
was) predisposed by years of happy experiences to "like" and "trust"
Suns or are irritated by the technical flaws in Intel CPUs. I don't
have to "like" Intel or Intel CPUs if they are cheaper by a factor of
two or three and drive a perfectly functional system.
Obviously I except things like 10 processor systems and speciality
systems from this -- I'm referring to the "commodity" workstation
business. Sun should well know, however, that the commodity workstation
business has a way of sneaking up on those high end high margin systems
and eventually eliminating them. Sun did just that to the VAX and the
IBM mainframe (with help, of course).
So, what are the lessons to be learned from this, and why does it show
that it is historically inevitable that linux and freebsd and their
cousins will (as Linux drolly puts it) achieve "world domination"?
a) Linux has filled the OS niche that COULD have been filled by Sun
back in 1988 or Digital in 1993 or Novell in 1995 (boy did Novell blow
THAT one!). The $50 (or less!) distribution is finally here and has
achieved critical mass (the number of systems required for
entrepreneurial developers and real corps to start to write real
software for it). It is in the phase of exponential growth, driven by
price/performance alone, into an "infinite" potential market. At this
point even if Sun or Novell "saw the light" and dropped prices to match,
they couldn't catch up. Microsoft, of course, could still win. All MS
has to do to win is drop prices for NT-server with all compilers and
programming and networking tools included to $50, retire Windows
95/98/2001 (and give people FREE upgrades to NT server from any MS OS
product) and maybe bundle NT with MS Office 97 for personal use only.
So Microsoft is doomed to lose.
b) Linux (more than any other OS ever) fully recognizes the difference
between "kernel", "libraries" and "distribution". The kernel is really
irrelevant to most users EXCEPT that it be functional and invisible and
stable. Libraries and posix compliance and the like matter only to
programmers and technoweenies. Users (and perhaps most systems
administrators who aren't also systems programmers) care only about
what's in a distribution. You like BSD? Well, there is freebsd or
there is slackware. You like SysV? Well, Red Hat is very SysV-ish in
its /etc/layout and management. Like Windoze-like tools? Some linux
distributions come with GUI system management tools. Like to do it
yourself? Some distributions are very flatfile oriented and presume
that either you know what you are doing or want to learn. Don't like
any of the distributions? Make your own. Do a good job, and it could
even become popular and displace the existing ones. Everybody likes
choices, and linux has more choices than a chinese restaurant, from
any-num-num-can-do-it Red Hat to build an entire distribution yourself
from source and lay it out the exact way you like it (Unix Guru's Only,
of course). Run it on a personal box at home (I do) or run an entire
network of it at work (I do, or rather I help).
c) Linux runs on true commodity hardware. Intel, AMD, Cyrix, Digital,
Sparc... heck, linux runs on NON-commodity hardware from other Unix
workstation vendors and not infrequently runs as well or better than the
vendor's own pet flavor of Unix. I could be wrong (NT runs on many
platforms as well) but I >>think<< that linux runs on more hardware
platforms than any other OS. Again, people love choices. Sun didn't
understand that porting SunOS to the x86 family of systems ultimately
would make people MORE likely to convert to Sparc boxes, not less, as at
the time Sparc systems tremendously outperformed any Intel system and
running the same OS on both would make it painless to jump from an x86
to a Sparc when more performance was required. Now users of linux would
hardly be inconvenienced if Intel main CPU factory were hit by a small
asteroid. They'd simply buy something else (alphas? Power PC's?) as
their old Intel boxes aged out, expending a lot of effort in a short
amount of time to debug the distributions as numbers increased.
...and then there is Gnome. Microsoft's last refuge is their Windows
interface (and all the software written for it), warts and all. As the
above clearly demonstrates, there is a huge real cost in switching user
or administrative interfaces. Over time, huge dollar differentials in
cost/benefit still motivate the change, but a Windows user or NT user is
probably lazy and fearful (just like me) and hesitates to try something
different if it will require a long time to learn to use it or even set
it up. The "one place" that many linux distributions could still be
simpler (including Red Hat) is in the very lowest end of userspace, the
Mac-loving dummy. Gnome is significant not so much because it is
supposed to be a super-duper GUI/WM in its own right but because it MAY
finally move num-num-user account management and userspace GUI
management down to the level that the average Mac or Windows user can
handle it.
In the case of Windows, the growing complexity of Windows is a point in
favor already -- if Windows DIDN'T come pre-installed on most systems,
few users would be able to handle the installation any more. I say this
with certain knowledge from the many times I've had to (re)install it on
various boxes I own. The mac still has the advantage of totally
protected and homogeneous hardware support so users can usually manage
an install, but Gnome promises to challenge even that. Give linux one
more year of (well-funded!) development toward maturity and any num-num
will be able to pop a floppy in the drive, a CD in the other drive,
boot, and manage a full install through account creation, add-on
software installation, and GUI setup and management all without opening
a manual. Red Hat is not TOO far from this now, although I think that
one is still way to likely to need to know what one is doing to make
things work correctly. This is not intended to invite a religious war
on distributions; the beauty of linux is that there is a CHOICE of
distributions and the open opportunity to roll your own. In three
years, when Microsoft finally abandons Windows as a primary moneymaker
and enters the linux reseller market to avoid bankruptcy, we could see
intense competition between linux distribution resellers (to the benefit
of all consumers).
The new world cometh. Linus Torvalds will yet stand on his mountain,
gentle ruler of all that he surveys...
rgb
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:[EMAIL PROTECTED]
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
