[rescue] Sunburst

[Skeezics Boondoggle]

On Thu, 20 May 2010, Andrew Hay wrote:

> " From: Nemo <cym224 at gmail.com>
> "
> " Here is a story of a wholly subjective comparison of Sparc versus
> " Intel.  I had a U10 (440 MHz and 1 GiB) running Sol 10 as well as an
> " IBM PC (~500 MHz and 1 GiB) running SuSE under my desk.  I was
> " compiling a patched gcc 4.1.2 (for a language that used the gcc
> " front-end) and needed to compile a patched gdb as well.  Both were
> " compiling on the U10 and I was editing with emacs as they trudged
> " along.  I thought: Why not try that on the PC?  The dual compilation
> " brought the PC to its knees with a noticeable delay between the typing
> " of a character and its appearance on the screen.  Compilation on the
> " U10 finished long, long before that on the PC.  (I know. I know.  The
> " PC architecture is crippled, still...)
>
> i'd venture to surmise that s10 is way more polished at multitasking
> than linux.  s10 x86 vs. linux would be illuminating...

Actually, I had an opportunity to do exactly that: test and benchmark
two identically configured Sun V20zs several years back, when S10
11/06 and SLES9 were current.  On identical hardware Solaris 10 was
8%-32% faster on every test we cared about (NFS, MySQL, Apache/Tomcat
+ Java, bunch of company-specific stuff).  Even a full custom
Jumpstart was faster than an equivalent AutoYaST, which was surprising
(or not, if you've ever been subjected to AutoYaST :-).  Really, like
15 _minutes_ faster, including setting up mirrored boot & root,
Cfengine localizations, etc.  That kinda blew my mind.

Ten years ago I had a 200Mhz Ultra-1 outperform a 500Mhz P-III doing
multiple compiles (Solaris 7 on both hosts, Sun's WorkShop 6
compilers) in pretty much the same scenario as the original poster.  I
was building up all the usual GNU/Apache/open source libs and packages
for a new dev/prod environment for a startup, y'know, back in the bad
old days before Sun even acknowledged that Perl existed. :-)  The U1
was running my OpenWin desktop/editor/xterms and acting as our
internal DNS, NIS, NTP, and DHCP/boot server, while the P-III was only
doing compiles.  Building one thing at time, the x86 box was, as you'd
expect, quite a bit faster.  Run two builds in parallel and it'd lag
pitifully, while the SPARC cruised right along finishing in step with
or even ahead of the x86.

But that has always been my experience with Intel boxes.  I've seen
100+ interactive users working away comfortably on a SS5, when a then
"beefy" Pentium running the same job mix would fall over dead with 20
or 30 users.  I've seen this repeated through multiple generations,
from the older SuperSPARCs to the US-IV+: they degrade gracefully
under extremely heavy loads, while x86 boxes fall over badly.
Normalized for clock rate, the SPARCs actually _could_ usually compete
with the Intel chips on single-threaded tasks, but if you could get
2-3x the clock rate at 1/2-1/3 the sticker price, it was always an
extremely hard sell the get management to believe that the
cheaper/"faster" box wasn't the best decision.  But now, with Oracle
apparently doing everything they can to make Solaris and SPARC even
more irrelevant and obscure than they already were, it's likely moot
anyway.  Soon not many people outside of lists like this will remember
or even care about this stuff.

OBTW,  that U1E I ran all those builds on ten years ago is still
running (albeit with a dead NVRAM battery!) as a
console/DNS/NIS/NTP/paging server here at home:

SunOS headbone 5.7 Generic_106541-44 sun4u sparc SUNW,Ultra-1
  3:01am  up 1147 day(s), 5 min(s),  3 users,  load average: 0.10, 0.06, 0.02

-- Chris