****** NEW SparcServer 1000 ******

****** NEW SparcServer 1000 ******

Post by Christian Marcot » Fri, 07 May 1993 13:27:21



I just read some gossip about the new sparcServer 1000 (code named Scorpion).
It's supposed to be released May 11th.
It's supposed to be a couple of pizza box high with the same footprint.
It's an 8 way multi-processor (first release with 50MHz SuperSparc+),
and up to 2Gb of memory with 200Mb/second I/O bandwidth, 12 expansion slots,
100Gb disk capacity and built-in high-availability fault-resilient commercial
features (From what my sun-type told me, I think they are talking about
their new feature where in case of fault, the hardware reboots automatically
and reconfigures itself around (and without) the faulty component (whoooo
that sounds like a mainframe :->).

It's cheap too, $36,700 for a uniprocessor with 32M ram & 1G disk,
$46,700 for a four way with 64M ram & 2Gb disk, $75K for 4way, 128Meg ram &
2G disk ...

It's supposed to top at 9000K+ specInt92 (and is supposed to be somewhat
linear in performance increase per module which makes it about 1000 specInt92
per CPU module !!!????

With a box like that, who can still say that Sun is out of the performance
race?

However one thing looks strange,  How do they get 1000 specInt92 out of a
CPU module when they performances of Viking are reportedly sooo bad???

Now that I think of it, these numbers are somewhat similar to what was
claimed for the sparcServer 2000 (2 CPU = 2000 specInt92, 8 CPU = 8000 ...)

Anyone has experience with SparcServer 2000?
Anyone has opinions on these figure or would like to share some knowledge
of this parallel architecture (good and bad).

OH, BTW, Sun will claim that this is the fastest machine around and the
best bang for the buck...

-- Scoob

--
---------------------------------------------------------------------------
Christian Marcotte                Bell-Northern Research

Telephone: (613) 763-2782         Nepean, Ont., Canada K1Y 4H7

 
 
 

****** NEW SparcServer 1000 ******

Post by Mark Culot » Fri, 07 May 1993 18:59:39


:>
:> It's supposed to top at 9000K+ specInt92 (and is supposed to be somewhat
:> linear in performance increase per module which makes it about 1000 specInt92
:> per CPU module !!!????
:>

This is something that really yanks my chain.

The performance figures for Sun's multiprocessor servers is quoted in SPECrate,
NOT SPECmarks, and the two metrics have entirely different scales.

I haven't seen anyone else's multiprocessor boxes performance measured in
SPECrate.  Are there any examples?

Mark

--
------------------------------------------------------------------------------


        Software Engineer             |       uunet!centerline!mkc
        CenterLine Software Inc.      |
        10 Fawcett St.                |       (617) 498-3311 (voice)
        Cambridge, MA 02138-1110      |       (617) 868-6655 (FAX)

------------------------------------------------------------------------------

 
 
 

****** NEW SparcServer 1000 ******

Post by Todd Bernha » Sat, 08 May 1993 01:51:09



>I haven't seen anyone else's multiprocessor boxes performance measured in
>SPECrate.  Are there any examples?

Probably because their numbers aren't very impressive...but I'd be happy
to pass those numbers on to you...  Enjoy...

                        SPECrate_int92  SPECrate_fp92   AIM-3 (jobs/min)
                        ==============  =============   ================
 Sun SPARCcenter 2000   8,047(8-way)    10,600 (8-way)  1,631.2(8-way)
 HP 890S                4,301 (4-way)   4,685 (4-way)   NA
 IBM 980                1,404           2,960           NA
 Pyramid MIS-24ES       NA              NA              1,112(8-way)
 Sequent 2000/750       NA              NA              1,455(12-way)

Quote:>Mark

---todd

--
Todd Bernhard, Evangelist, Product Marketing,
Sun Microsystems Computer Corporation

 
 
 

****** NEW SparcServer 1000 ******

Post by Todd Bernha » Sat, 08 May 1993 04:37:15




>>                        SPECrate_int92  SPECrate_fp92   AIM-3 (jobs/min)
>>                        ==============  =============   ================
>> Sun SPARCcenter 2000   8,047(8-way)    10,600 (8-way)  1,631.2(8-way)
>> HP 890S                4,301 (4-way)   4,685 (4-way)   NA
>> IBM 980                1,404           2,960           NA

>Lets see, isn't specRATE the marketoid's kludge of *linearly* adding
>the specINTs (or specFP) of all the processors together to get a very
>meningless number?  

No, it's not.

SPECrate is a specific benchmark.  It measures running multiple jobs
simultaneously, exercising all the CPUs of an MP (or UP) system).  

It measures how many tasks can be done each hour.  Hardly a 1:1 or 1:8
multiplication of SPECint/SPECfp...it is a real, specific benchmark.

Quote:>If we divide the above numbers by the number of processors we see that
>Sun is not out in front - but behind.

If your assumption was correct about SPECrate, than perhaps, but your
assumption is plain wrong.  Even if you were correct, in overall
performance, Sun leads.  Also, those numbers are with 40MHz SuperSPARC
parts.

Quote:>-wolfgang

---todd

--
Todd Bernhard, Evangelist, Product Marketing,
Sun Microsystems Computer Corporation

 
 
 

****** NEW SparcServer 1000 ******

Post by John R. Mash » Sat, 08 May 1993 06:28:36


|>                   SPECrate_int92  SPECrate_fp92   AIM-3 (jobs/min)
|>                   ==============  =============   ================

|>Sun SPARCcenter 2000        8,047(8-way)   10,600 (8-way)  1,631.2(8-way)
|>HP 890S             4,301 (4-way)  4,685 (4-way)   NA
|>IBM 980             1,404          2,960           NA
|>Pyramid MIS-24ES   NA              NA              1,112(8-way)
|>Sequent 2000/750   NA              NA              1,455(12-way)

SGI Challenge L, 4-way   5,562          6,131           N/A            
SGI Challenge L, 8-way  10,177          12,020          N/A
SGI Challenge L, 12-way 13,406          17,370          N/A

boards.

DEC also has some fairly high SPECrate_fp92s; I haven't seen SPECrate_int92s
for SMPs yet.

-john mashey    DISCLAIMER: <generic disclaimer, I speak for me only, etc>

DDD:    415-390-3090    FAX: 415-967-8496
USPS:   Silicon Graphics 7U-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311

 
 
 

****** NEW SparcServer 1000 ******

Post by Walter Ba » Sat, 08 May 1993 06:55:29


The SPECrate metrics were defined by SPEC specifically to preclude
marketeers from adding together SPECint92 or SPECfp92.  The scale was
set much higher than the speed metrics (e.g., an SS-10/51 has SPECint92
of 65.2 and a SPECrate_int92 of 1547) as one more way of ensuring that
nobody could confuse the two metrics.  You have to decide the relative
importance of speed versus throughput for your own workload.

The December 1992 SPEC Newsletter summarizes 57 SPECrate results
published for uniprocessors and multiprocessors from Control Data, Data
General, HP, Siemens-Nixdorf, Solbourne, and Sun.  The March 1993
newsletter adds new results for DEC Alphas, HP 755, IBM 370/375,
Siemens-Nixdorf RM600/25, Solbourne 6/900, and Sun SC-2000.  (*ad below)

*For newsletters, call NCGA at (703)698-9600 ext 318.  (U.S.)

---

Sun Microsystems,  2550 Garcia Ave.,  MTV15-404,  Mountain View,  CA   94043
(415)336-3689   SPEC Open Systems Steering Committee Chair FAX (415)968-4873

 
 
 

****** NEW SparcServer 1000 ******

Post by Robert Fara » Sat, 08 May 1993 10:18:27



Quote:>|>                     SPECrate_int92  SPECrate_fp92   AIM-3 (jobs/min)
>|>                     ==============  =============   ================
>|>Sun SPARCcenter 2000  8,047(8-way)   10,600 (8-way)  1,631.2(8-way)
>|>HP 890S               4,301 (4-way)  4,685 (4-way)   NA
>|>IBM 980               1,404          2,960           NA
>|>Pyramid MIS-24ES     NA              NA              1,112(8-way)
>|>Sequent 2000/750     NA              NA              1,455(12-way)
>SGI Challenge L, 4-way   5,562         6,131           N/A
>SGI Challenge L, 8-way  10,177         12,020          N/A
>SGI Challenge L, 12-way 13,406         17,370          N/A

Some *old* (Nov 92) figures I have for the 200 MHz AXP running VMS

DEC 10000-640, 4-way     9,108          17,187

cheers,
--

Networks Engineering            Disclaimer: Nothing I say is official, but
Digital Equipment Corporation    I will say "I hate it when you buy batteries
Gold Coast, Australia            and they're not included".

 
 
 

****** NEW SparcServer 1000 ******

Post by John - Madis » Sat, 08 May 1993 04:22:01


This looks like another example of benchmark bs from Sun.  Sun has fallen
into the habit of using unreleased hardware, unreleased software,
third party software not shipped with their machine, simulations, bogus
assumptions, and any other trick they can find to hide the fact that their
machines just don't stack up.  Witness their claimed 60-odd mip rating
for the Classic, or their linear specmark increase with number of processors.

I have been told by Sun employees that:
        - solaris mp performance needs a *lot* of work.  Forget about linear
          speed increases -- they were talking about total throughput
          remaining *constant* as the number of processors went up.
        - that the sparccenter 2000 mp numbers quoted by sun in their
          literature were fabricated.  At the time, Sun had run no mp
          benchmarks at all!

How about a description of the test setup and some independent verification
of these numbers?  Considering the abysmal performance of viking, it's
pretty hard to believe that their 8 processor machine is faster than
everyone else's.

>>I haven't seen anyone else's multiprocessor boxes performance measured in
>>SPECrate.  Are there any examples?

>Probably because their numbers aren't very impressive...but I'd be happy
>to pass those numbers on to you...  Enjoy...

> [suspect results deleted]

>--
>Todd Bernhard, Evangelist, Product Marketing,
>Sun Microsystems Computer Corporation


 
 
 

****** NEW SparcServer 1000 ******

Post by Todd Bernha » Sun, 09 May 1993 02:45:22



Quote:>I have been told by Sun employees that:
>    - solaris mp performance needs a *lot* of work.  Forget about linear

That's been improved greatly between Solaris 2.0, 2.1 and now 2.2.  You
may have an old impression...have you used Solaris 2.2?

Quote:>    - that the sparccenter 2000 mp numbers quoted by sun in their
>      literature were fabricated.  At the time, Sun had run no mp
>      benchmarks at all!

Wrong.  In fact, the email I forwarded was from October 22nd, before
the Nov 10th launch.  

Quote:>How about a description of the test setup and some independent verification
>of these numbers?  Considering the abysmal performance of viking, it's
>pretty hard to believe that their 8 processor machine is faster than
>everyone else's.

Consider that the SPARCcenter 2000 has a packet-switched bus (actually two
of them) that achieves 500MB/sec *sustained* bandwidth.  Competitors use
busses as slow as 128MB/sec or even 80MB/sec.  Much more goes into MP
performance than just CPU speed.

Also, note that the CPUs were 40MHz.  If you follow the news, we just
introduced 50MHz...Viking is doing just fine.

---todd

--
Todd Bernhard, Evangelist, Product Marketing,
Sun Microsystems Computer Corporation

 
 
 

****** NEW SparcServer 1000 ******

Post by Rick Jon » Sun, 09 May 1993 03:25:20


:
: It's supposed to top at 9000K+ specInt92 (and is supposed to be somewhat
: linear in performance increase per module which makes it about 1000 specInt92
: per CPU module !!!????
:

It is possible you are either confusing SPECrate_int92 with SPECint92,
or being fed mis-information.

Repeat after SPEC - SPECrate_int92 != SPECint92...

SPECrate_[int|fp]92 metrics are derived by running *multiple*,
*independent* copies of the component SPEC[int|fp]92 benchmarks. The
rate metrics are capacity measures and not speed metrics. For more
information about SPEC benchmarks, I would suggest that you poke
around comp.benchamrks. You might even find results for other
multiprocessor systems there.

Of course, you can always subscribe to the SPEC newletter - again, see
comp.benchmarks for more information.

rick jones
some of this could be wrong - I really only deal with SPEC SFS (a.k.a
LADDIS, the nhfsstone replacement, ask for it by name :)

 
 
 

****** NEW SparcServer 1000 ******

Post by John R. Mash » Sun, 09 May 1993 05:26:15


|> >|>                     SPECrate_int92  SPECrate_fp92   AIM-3 (jobs/min)
|> >|>                     ==============  =============   ================
                                                                        $ Approx        
|> Sun SPARCcenter 2000  8,047(8-way)   10,600 (8-way)  1,631.2(8-way)       ~$210K
|> >SGI Challenge L, 4-way   5,562         6,131           N/A            ~$120K
|> >SGI Challenge L, 8-way  10,177         12,020          N/A            ~$200K
|> >SGI Challenge L, 12-way 13,406         17,370          N/A            ~$280K
|>
|> Some *old* (Nov 92) figures I have for the 200 MHz AXP running VMS
|>
|> DEC 10000-640, 4-way     9,108          17,187            $320K + (3 CPUs)

I've added a couple of prices (approximate: tell me if I'm wrong).

Just out of curiosity, are there any published prices for the
MP versions of the 4000/7000/10000? i.e., incremental price /
CPU board?),
(I haven't happened to see any so far;
uniprocessor list prices, I think are about US$80K, $170K, $320).

For the Challenge machines, another $50K for L (deskside) ==> XL (rackmount);
the prices are actually for the (forthcoming, faster) 150Mhz versions.

-john mashey    DISCLAIMER: <generic disclaimer, I speak for me only, etc>

DDD:    415-390-3090    FAX: 415-967-8496
USPS:   Silicon Graphics 7U-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311

 
 
 

****** NEW SparcServer 1000 ******

Post by John R Mu » Sun, 09 May 1993 09:07:55



>That's been improved greatly between Solaris 2.0, 2.1 and now 2.2.  You
>may have an old impression...have you used Solaris 2.2?

That might be kind of hard, Todd, unless I've slept through a couple of
weeks.  From a recent Sun news release:

        Solaris 2.2. will be available at the end of May, 1993 and will
        become the default release for all Sun SuperSPARC-based multi-
        processing systems.

Folks, could we either tone the marketing hype down just a *little* bit
here, or take it elsewhere, like /dev/null?  Thank you.

... John