Very Computer

(1) Sun-SunOS, Solaris,
(2) IBM PPCs - AIX, Solaris ppc
(3) Pentium P5/P6 - BSD (I, net), x86 Solaris

After looking at all the respective price/performance ratios I
concluded that the only efficient way to go is with the P6-based
machines.  I would like to go with Sun and SunOS or Solaris, but these
machines are the worst in price/performace ratios ($/perf.).  These
$/perf. ratios for the Intel based machines are MUCH better than that
for the PPC or Sun platforms.  I do not like any company that has close
ties with Microsoft, but from a business point of view, I can not
ignore them any longer.  The reason for this post is that I?m much
confused by this issue.  Many people prefer Sparcs (maybe I would as
well), but business decision should be based more than on a personal
preference alone.  Please read what I did, fell free to make comments,
and if you can or are willing to, answer my questions below.

Any ideas are welcomed....thank you for your time, Martin Stary

----------------------------------
Method:  I have compared the three different platforms on price using
specific manufacture?s system prices, and performance using several
SPEC benchmarks.  I know that the SPEC benchmarks are not truly
indicative of the overall system performance, but from some of my
experience the CPU benchmarks do correlate with the performance of the
machine as a whole.  For pricing, I have used several machines in each
platform group configured in a similar way.  There are some
discrepancies in the configurations, but these will most likely not
amount to significant price differences.  For the Intel based systems,
the price is somewhat higher than they really are, and for the PPC and
Sparc based systems the prices are representative or lower than they
should be (even this does not help the Sparcs and PPCs).

Conclusions:  I can not justify to buy any other platform other then
the Intel P6/200 based machines.  The $/perf. ratios (look below) for
the P6/200s are in many instances twice as good as the next viable
platform (Sparc-based).  The P6/200 boxes seem to have a HUGE advantage
in $/perf.  If this advantage was only 20-30% different then I could
overlook the P6, but not when the machines (P6) seem to have twice (in
many instance way more than twice) the horsepower per $.  The
UltraSparcs seem to have better $/perf. ratios compared to the older
Super/Hyper Sun models, but not anywhere near the P6/200 machines.
Since SunOS is not supported by the new UltraSparcs, then I may as well
run Solaris on a P6/200 significantly cheaper.

Questions:
(1) Am I correct for the most part, or am I making some fatal mistakes?
 If I?m making mistakes, where are they and how can I correct them?
(2) What would be a better measure of overall ISP/WWW server
performance then those used here?  There is the measure of tpm
(transaction per minute, or something similar), but I can not seem to
find this data for these specific machines.

Data:
I found a ?Spectable? somewhere on the net, but I do not remember
where.  I hope that I?m not infringing on the original author?s rights.

All of the systems are configured with 64Megs of RAM, 4.3 gig Hds, some
graphics card (not very important for servers), and no monitors.

Here are the data that I?m deriving my conclusions from:
-----------------------------------------
SPECint_base95, SPECfp_base95
-----------------------------------------

Machine, CPU, clock, cache, SPECint_base95, SPECfp_base95, price

Sun SS10/40       SuprSP     40      20/16        1.00    1.00    ???
IBM C20           MPC604     120     1M+16/16     3.38    3.48  
$12300
Intel XXpress     Pentium    66/133  1M+8/8       4.14    3.12  
~$6000
Intel Alder       PentiumPro 150     256+8/8      6.08    5.42  
~$6700
Intel Alder       PentiumPro 200     256+8/8      8.09    6.70   ~$7200
Sun SS20/71       SuprSP2    75      1M+20/16     2.06    2.14  
$10400

$/perf. ratios for the above:
CPU     $/int_base95    $/fp_base95
--------------------------------------------
MPC604   $  3,639.05     $  3,534.48
P5/133   $  1,449.28     $  1,923.08
P6/150   $  1,101.97     $  1,236.16
P6/200   $      889.99   $  1,074.63
SuprSP2/71       $  5,048.54     $  4,859.81

========================
SPECint92, SPECfp92
-----------------------------------------

Machine, CPU, clock, cache, SPECint92, SPECfp92, price

DEC VAX11/780     VAX        5       2             1.0     1.0    ????
IBM C20           MPC604     120     1M+16/16    155.0   150.2  
$12300
Sun SS5/110       MicroSP2   110     16/8         78.6    65.3    $8800
Sun SS20/50       SuprSP     50      20/16        76.9    80.1    $7700
Sun SS20/61       SuprSP     50/60   1M+20/16     98.2   107.2    $9300
Sun SS20/71       SuprSP2    50/75   1M+20/16    125.8   121.2  
$10400
Sun SS20/612      2xSuprSP   50/60   1M+20/16      ?     127.1  
$12100
Sun Ultra1/140    UltraSP    71/143  512+16/16   215     303    ~$14000
Sun Ultra1/170    UltraSP    83/167  512+16/16   252     351    ~$15000
Intel Xpress      Pentium    66/133  512+8/8     147.5   109.6  
~$6000
Intel Alder       PentiumPro 150     256+8/8     276.3   220.0  
~$6700
Intel Alder       PentiumPro 200     256+8/8     366.0   283.2  
~$7200

$/perf. ratios for the above:
CPU     $/SPECint92      $/SPECfp92
--------------------------------------------
MPC604/120      $79.35  $81.89
MicrSP2/110      $111.96        $134.76
SuprSP/50       $100.13 $96.13
SuprSP/61       $94.70  $86.75
SuprSP2/75      $82.67  $85.81
2SuprSP2/60     ???     $95.20
UtlraSP/143     $65.12  $46.20
UltraSP/167     $59.52  $42.74
P5/133  $40.68  $54.74
P6/150  $24.25  $30.45
P6/200  $19.67  $25.42

========================
Integer/FP SPECrate92
-----------------------------------------

Notes:
        - Integer SPECrate is derived from the results of a set of
integer
          benchmarks run multiple times simultaneously, and can be used
          to estimate a machine's overall multi-tasking throughput for
          integer code. It is typically used on MP machines.
        - Floating-Point SPECrate is derived from the results of a set
          of floating-point benchmarks run multiple times
simultaneously,
          and can be used to estimate a machine's overall multi-tasking
          throughput for FP code. It is typically used on MP machines.
        - Computed specrates are indicated by "c". They're computed
from
          SPECint92, SPECfp92 (for uniprocessors) using a scaling
factor.
          This number is usually slightly less than or equal to a
measured
          specrate on a uniprocessor. The scaling factor is the number
of
          seconds in a week, divided by the time of the longest-running
          benchmark on the reference SPEC VAX 11/780, which is
604800/25500,
          or about 23.7.

IBM C20           MPC604     120     1M+16/16     3676c   3562c  
$12300
Sun SS5/110       MicroSP2   110     16/8         1864c   1549c  
$8800
Sun SS20/61       SuprSP     50/60   1M+20/16     2092    2418    $9300
Sun SS20/502      2xSuprSP   50      20/16        3218    3193    $8400
Sun SS20/612      2xSuprSP   50/60   1M+20/16     4492    4888  
$12100
Sun SS20/712      2xSuprSP2  50/75   1M+20/16     5726    5439  
$14500
Sun SS20/HS14     4xHyperSP  50/100  256+8/0      8124    8906  
$19500
Sun Ultra1/140    UltraSP    71/143  512+16/16    5107    7175  
$14000
Sun Ultra1/170    UltraSP    83/167  512+16/16    5982    8323  
$15000
Intel Xpress      Pentium    66/133  512+8/8      3498c   2599c  
$6000
Intel Alder       PentiumPro 150     256+8/8      6553c   5218c  
$6700
Intel Alder       PentiumPro 200     256+8/8      8681c   6717c  
$7200

$/perf. ratios for the above:
CPU     $/Integer/FP    $/SPECrate92
--------------------------------------------
MPC604/120      $3.346  $3.453
MicrSP2/110     $4.721  $5.681
SuprSP/61       $4.446  $3.846
2SuprSP/50      $2.610  $2.631
2SuprSP/612     $2.694  $2.475
SuprSP2/712     $2.532  $2.666
4HyprSP/100     $2.400  $2.190
UtlraSP/143     $2.741  $1.951
UltraSP/167     $2.508  $1.802
P5/133  $1.715  $2.309
P6/150  $1.022  $1.284
P6/200  $0.829  $1.072

Powerstar Inc.                                                  SCSI-3,
UPS Technologies
800-209-5556

Quote:> All of the systems are configured with 64Megs of RAM, 4.3 gig Hds, some
> graphics card (not very important for servers), and no monitors.

In any case, for a web server, I'd consider a 43P workstation rather than
the C20 server.  The base price is well less than half for a PPC 604/133-
equipped version, and  it includes networking in the box.  I'm not sure
how you arrived at your prices, so I'll guesstimate below.

Quote:> -----------------------------------------
> SPECint_base95, SPECfp_base95
> -----------------------------------------

> Machine, CPU, clock, cache, SPECint_base95, SPECfp_base95, price

> Sun SS10/40       SuprSP     40      20/16        1.00    1.00    ???
> IBM C20           MPC604     120     1M+16/16     3.38    3.48   $12300
> Intel XXpress     Pentium    66/133  1M+8/8       4.14    3.12   ~$6000
> Intel Alder       PentiumPro 150     256+8/8      6.08    5.42   ~$6700
> Intel Alder       PentiumPro 200     256+8/8      8.09    6.70   ~$7200
> Sun SS20/71       SuprSP2    75      1M+20/16     2.06    2.14   $10400

Quote:> $/perf. ratios for the above:
> CPU        $/int_base95    $/fp_base95
> --------------------------------------------
> MPC604      $  3,639.05     $  3,534.48
> P5/133      $  1,449.28     $  1,923.08
> P6/150      $  1,101.97     $  1,236.16
> P6/200      $    889.99     $  1,074.63
> SuprSP2/71  $  5,048.54     $  4,859.81

Based on those, the 43P is in the same ballpark as the Pentium 133.  But
you'll have to work out exact numbers for yourself to see whether it's
really comparable to Intel boxes.

--

[ "any sufficiently advanced technology                                      ]
[  is indistinguishable from magic"                      -- arthur c. clarke ]

Do you want to be fastest or do you want to stay up the longest?

People are discovering the problems with the P6 right now.  Last
I heard was a horrible IO choke-point that won't get solved until
a revision of the P6 MB.

I've got P100s cruising fast with all of the PCI bus I need and
they stay up for 50 days at a time if I need them to. They are
running on the proven Neptune II chip-set with the only PCI
disk controllers that are "approved" by BSDI (Buslogic).  And
I spend Zero time tending to them.  When I need more "performance"
I either add memory, add disks (maybe a controller), or add another
system.  I don't even worry about auto-rebooter's by phone and the
like.  My life is calm :)

I've got friends who are doing Unixware Dual-Pentium on the Triton
Chipset and all the other fancy stuff (new Adaptec cards).  That
seems to mean unpredictable crashes and other headaches.  Customers
don't like it at all......

BSDI is going to run around SunOS, Solaris, and AIX .  It is faaaaast.
The Sparc20s are very nice and some of them might actually be an
improvement -- but why move yourself off of the commodity prices of
Intel sub-systems ?

If this is just your play money, work with your favorites that
make you happy.  

If this is the company money, think about what you can do to
give them (and yourself) the least headache -- since your time
is far more expensive than the equipment or software.

Just my stenkin' opinion....

--

"He who loses control, loses."                    http://www.veryComputer.com/
Frank Pembleton, *

Quote:

>Hi,
>I have a dilemma about the hardware that I should use for my ISP/WWW
>site (www, ftp, ppp/slip, news, pop....).  After some research and
>after reading the Inet-access FAQ a couple of times, I get very
>contradicting evidence on what h/w I should use.  I have considered and
>analyzed three different platforms (from a raw CPU performance point of
>view):

>(1) Sun-SunOS, Solaris,
>(2) IBM PPCs - AIX, Solaris ppc
>(3) Pentium P5/P6 - BSD (I, net), x86 Solaris

Although Sun's hardware has the ability to multitask MUCH better than
any Intel processor, their OS sucks rocks for performance.  

Intel is definately the cheapest way to go, and if you (as an ISP) aren't
going to do much processing, but lots of PPP sessions and routing, email,
etc., Intel is a great platform, especially with FreeBSD or BSDi.

If you want a web server which runs LOTS of PERL scripts, does database
searches, etc., then maybe you need to check out Silicon Graphics.
These platforms CRANK.

Advise: get a reseller's license so you can get a price break on SGI
equipment, or buy an Indy used.

Good luck!

Quote:>BSDI is going to run around SunOS, Solaris, and AIX .  It is faaaaast.
>The Sparc20s are very nice and some of them might actually be an
>improvement -- but why move yourself off of the commodity prices of
>Intel sub-systems ?

Quote:>If this is the company money, think about what you can do to
>give them (and yourself) the least headache -- since your time
>is far more expensive than the equipment or software.

btw, those int/$ and fp/$ """benchmarks""" are totally bogus and deceptive -- in other
words, bullshit && waste-of-time (TM).

Chris

--___  ____ __
 | _ \/ __/|  \   Christian Kuhtz, Pencom System Administration Services

 |_|  /___/|_|__\ on-site at IBM Government Systems, Boulder, Colorado

It takes about 30s to recognize the controller.  Aggregate throughput
through the file system exceeds 16 MB/s.  Costs are ~PC mail order.
Reliability is becoming quite well known.

Best regards,
Russell Carter
http://www.geli.com  Pentium Workstation Clusters

Quote:>Method:  I have compared the three different platforms on price using
>specific manufacturers' system prices

Quote:>SPEC benchmarks.  I know that the SPEC benchmarks are not truly
>indicative of the overall system performance, but from some of my
>experience the CPU benchmarks do correlate with the performance of the
>machine as a whole.

Quote:>many instance way more than twice) the horsepower per $.  The
>UltraSparcs seem to have better $/perf. ratios compared to the older
>Super/Hyper Sun models, but not anywhere near the P6/200 machines.

Quote:>Since SunOS is not supported by the new UltraSparcs

Quote:>then I may as well run Solaris on a P6/200 significantly cheaper.

Quote:>All of the systems are configured with 64Megs of RAM, 4.3 gig Hds, some
>graphics card (not very important for servers), and no monitors.
...
>(1) Am I correct for the most part, or am I making some fatal mistakes?

o You only consider SPARC machines from Sun.  You might find that considering
  machines from Axil, Tatung, or Integrix would change the cost significantly.

o You probably aren't considering discounts from Sun (and probabl IBM, too)

o You're throwing in framebuffers that are almost certainly not comparable.
  For a network-services machine, you don't need *any* graphics device for
  a Sun, at least.  The graphics hardware in Suns is generally a much different
  beast from the price-point cheap stuff in an MS-DOS-market machine.  The
  latter rarely can usably support a million pixels, and probably don't offer
  the speed and acceleration that the Sun card probably does.  As such, the
  Sun card is going going to cost more, especially list.  I suggest pricing
  a configuration without a framebuffer to be more fair.  I don't think you
  mentioned monitors, but if you're including a monitor with the Suns, then
  that's another mistake.  MS-DOS machines seem to rarely be sold with decent
  monitors.  I rarely see one as large as 17", and they're almost always
  spherical, and almost always can't handle even close to 1M pixels without
  flickering.  I don't think Sun sells anything smaller than 17" now, and I
  believe that they only sell cylindrical Sonys doing at least 1152x900.
  Again, this is a different beast from an MS-DOS monitor, so's it's gonna
  cost more.  You don't need one for a Sun, so don't price it in.

o You're probably pricing the Sun and IBM machines with memory and disks
  OEM'd from Sun and IBM, respectively.  Buying both elsewhere would probably
  save a good bit of $$.

I'm not saying that the totals will necessarily change so dramatically as to
be obviously in favor of the native-Unix platform, but if you're going to

Quote:>o You're throwing in framebuffers that are almost certainly not comparable.
>  ...The graphics hardware in Suns is generally a much different
>  beast from the price-point cheap stuff in an MS-DOS-market machine.  The
>  latter rarely can usably support a million pixels, and probably don't offer
>  the speed and acceleration that the Sun card probably does.

Acceleration is not that big a deal; X mostly wants a fast framebuffer
for day-to-day use, and a PCI card provides that. There are some
very fast and very expensive graphics cards available, and yes,
they may not be as good as Sun's top-of-the-line stuff. But for
those of us not doing real-time 3D modelling, there's no difference.

Quote:>  ...but if you're including a monitor with the Suns, then
>  that's another mistake.  MS-DOS machines seem to rarely be sold with decent
>  monitors.  I rarely see one as large as 17", and they're almost always
>  spherical, and almost always can't handle even close to 1M pixels without
>  flickering.  I don't think Sun sells anything smaller than 17" now, and I
>  believe that they only sell cylindrical Sonys doing at least 1152x900.
>  Again, this is a different beast from an MS-DOS monitor, so's it's gonna
>  cost more.  You don't need one for a Sun, so don't price it in.

cjs
--

Internet Portal Services, Inc.  
Vancouver, BC   (604) 257-9400          De gustibus, aut bene aut nihil.

  If the vendor, eg., Sun, won't give you the usual 20-40% discount

Quote:>off of list, then you can do at least as well by buying equivalent machines
>from somebody like Axil, Integrix, or Tatung.

Quote:>I've recently read that some (many? most?) of the MS-DOS-architecture machines
>can't cache more than, say, 64M or 128M of memory, and that memory above
>that will be uncached.  This is both amazing and scary.

Quote:>o You're throwing in framebuffers that are almost certainly not comparable.
>  For a network-services machine, you don't need *any* graphics device for
>  a Sun, at least.

Quote:> The graphics hardware in Suns is generally a much different
>  beast from the price-point cheap stuff in an MS-DOS-market machine.  The
>  latter rarely can usably support a million pixels, and probably don't offer
>  the speed and acceleration that the Sun card probably does.

But, as was stated fast graphics card are worthless in a server.

Quote:>  that's another mistake.  MS-DOS machines seem to rarely be sold with decent
>  monitors.  I rarely see one as large as 17", and they're almost always
>  spherical, and almost always can't handle even close to 1M pixels without
>  flickering.

PC's have come a long way since the original IBM PC.  Don't discount them
just because they have been running crappy OS's in the past.

Nate

--


work #: (406) 449-7662 | Montana.
home #: (406) 443-7063 | A fly pole and a 4x4 Chevy truck = Heaven on Earth

Quote:>In order to configure such a box to run a Unix decently, you don't
>use a flinky $25 mass-market enclosure -- you pay more for one with a decent
>power supply and cooling.  You pay more if you want a halfway-usable keyboard,
>for any machines that need one.  In the end, though, memory and disk costs
>for this sort of application are going to dwarf the CPU costs anyway, which
>kinda makes the argument moot.

Quote:>IMHO, the Unix-on-MSDOS-hardware route can be by far the most time-consuming
>and headache-prone.  Hardware and software vendor support can be even worse
>than that of Unix-hardware vendors.

Quote:>It can take a week to get a machine to simply recognize two SCSI
>controllers.  Nobody expects you to put two into a machine, so nobody
>tells you the convolutions needed to do so.

Quote:> Essential utilities like a
>disk analysis tool can be missing, and lots of tools may not build at
>all on the platform -- eg., lsof.

Nate
--


work #: (406) 449-7662 | Montana.
home #: (406) 443-7063 | A fly pole and a 4x4 Chevy truck = Heaven on Earth

Many people think more address bits is more speed but they've been
brainwashed by Intel and the x86 compilers that suck: most Intel
compilers can't/couldn't mix 16bit addressing with 32 bit arithmetic so
you needed to use 32 bit all the way to get the advantage of newer
instructions and thus the 32bitness seemed to be the performance
improver.  More databits matter, though, especially when going from
16->32.

The most important requirement when making UltraSPARC machine w/ 2.5
was: they must run all current SPARC apps unchanged.  And they seem to
be living up to that promise well.

Casper
--
Expressed in this posting are my opinions.  They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.