Very Computer

I'm facing an odd problem.

I recently build a new machine based around an AMD recommended motherboard
and K6-2/500.  The strange thing is that with two identical Linux
installations this machine appears to be a little less than 20% faster
than my P160 machine (overclocked at 200).  The test here is setiathome
which is getting 99%-plus CPU time on both machines (according to top).  
Watching the progress counts on the two machines side by side shows
that for every 10% completed on the old machine the new machine has
processed 11.7% or thereabouts.

What am I missing?  What do people recommend in the way of "proper"
benchmarks?  Can I log kernel stats and perform the equivalent of
sar reports on commercial platforms?

(I heard some rumblings about Linux not behaving well with more than
64Mb of RAM.  Since the new machine has 128Mb I wonder, is this a true
concern?)
--
Raj Rijhwani        (umtsb5/16) |  This is the voice of the Mysterons...

                                |  "Lieutenant Green:  Launch all Angels!"
http://www.courtfld.demon.co.uk/raj/ (demon, and gods, willing...)

What clock speed and memory speed does that K6-2/500 use?  The older K6
chips are not known for their fpu speed.  When I went from P180 MMX


times faster (~8 hrs) and compiles a kernel more than twice as fast.

Quote:>What am I missing?  What do people recommend in the way of "proper"
>benchmarks?  Can I log kernel stats and perform the equivalent of
>sar reports on commercial platforms?

Benchmarking-HOWTO

Quote:>(I heard some rumblings about Linux not behaving well with more than
>64Mb of RAM.  Since the new machine has 128Mb I wonder, is this a true
>concern?)

Note likely.  I can definitely tell you that 96 meg works better than 32
meg, since it just about eliminates swapping.

--

http://www.autox.chicago.il.us/  http://www.berniesfloral.net/
http://hammer.prohosting.com/~cgi-wiz/  http://cgi-help.virtualave.net/

Quote:> What clock speed and memory speed does that K6-2/500 use?  The older K6
> chips are not known for their fpu speed.  When I went from P180 MMX

Bus speed 100MHz, chip speed 500Mhz, fast memory in the BIOS, can't
remember the actual memory times.

Hmmm - maybe I should consider substituting a PIII...

Quote:> Benchmarking-HOWTO

Doh!  Why is it (with the complete HOWTOs on disk) I never think to
look there?  I must have a mental block.  Thanks.

Quote:> >(I heard some rumblings about Linux not behaving well with more than
> >64Mb of RAM.  Since the new machine has 128Mb I wonder, is this a true
> >concern?)
> Note likely.  I can definitely tell you that 96 meg works better than 32
> meg, since it just about eliminates swapping.

I thought this was probably a bogus argument when I heard it, but I
wondered.
--
Raj Rijhwani        (umtsb5/16) |  This is the voice of the Mysterons...

                                |  "Lieutenant Green:  Launch all Angels!"
http://www.courtfld.demon.co.uk/raj/ (demon, and gods, willing...)

Quote:> I recently build a new machine based around an AMD recommended motherboard
> and K6-2/500.  The strange thing is that with two identical Linux
> installations this machine appears to be a little less than 20% faster
> than my P160 machine (overclocked at 200).  The test here is setiathome

Seti is very FP and memory intensive, and isn't a representative test of
system performance. (Of course, it's a very good test if you purchased the
system to run setim but that presumably wasn't your primary intention). The
K6-2 has a relatively weak FPU. Try compiling identical kernels with
identical configs on each (using the same version of everything) and see what
happens. This should scale at linearly with CPU power, and more if the K6-2
machine has more RAM.

Quote:> (I heard some rumblings about Linux not behaving well with more than
> 64Mb of RAM.  Since the new machine has 128Mb I wonder, is this a true
> concern?)

I think this comes from old machines not having the ability to cache memory
above 64MB. Not a concern with modern hardware.

Charlie

There is a minor problem in that with some BIOSes Linux does not correct
discover memory above 64Mb, so it only uses the first 64.  If you have
that problem you just have to add an append argument to your lilo.conf
file to tell the kernel explicitly how much memory it has and all is
fine.

I have 128Mb on my work machine and my home machine.  The home machine
needed the fix but both work reliably with 128.

--

Voice: +44 1473 642987         MSMAIL:   BTEA/BTLIP23/FOSDICSJ
Fax:   +44 1473 646656         BOAT:     FOSDICSJ
Snail: B29/G34, BT Labs, Martlesham Heath, Ipswich, IP5 7RE, England.

Quote:>There is a minor problem in that with some BIOSes Linux does not correct
>discover memory above 64Mb, so it only uses the first 64.  If you have
>that problem you just have to add an append argument to your lilo.conf
>file to tell the kernel explicitly how much memory it has and all is
>fine.

You are confusing two distinct problems.

Older versions of the _kernel_ do not automagically discover memory above
64Mb, irrespective of the BIOS - this needs the append argument.

Older motherboard chipsets do not cache memory above 64Mb, meaning that
adding memory to these systems damages performance; there is little one
can do about this in software.
--

http://www.chiark.greenend.org.uk/~damerell/ w.sp.lic.#pi<largestprime>.2106
|___|          Any sufficiently technologically advanced music         |___|
| | |               is indistinguishable from line noise.              | | |

On 08 May 2000 17:10:55 +0100 (BST),

Why do you say that?

Quote:> Older versions of the _kernel_ do not automagically discover memory above
> 64Mb, irrespective of the BIOS - this needs the append argument.

And current versions with some newer BIOS's.  (Check deja for this going
back to December -- some BIOS makers changed the API they use around
then, and several people who 'upgraded' their BIOS found that even the
current kernels would no longer detect their RAM correctly.)    

Even with 2.2, some systems still require the 'mem='.

Quote:> Older motherboard chipsets do not cache memory above 64Mb, meaning that
> adding memory to these systems damages performance; there is little one
> can do about this in software.

I don't see why you think the followup was referring to this.

--
Brian Moore                       | Of course vi is God's editor.
      Sysadmin, C/Perl Hacker     | If He used Emacs, He'd still be waiting
      Usenet Vandal               |  for it to load on the seventh day.
      Netscum, Bane of Elves.

spake unto us, saying:

Very true.  My IBM IntelliStations are not new (PPros with 128MB and
96MB respectively), but the 2.2.14 kernel packaged with Mandrake 6.1
can't see anything at all about 64MB unless explicitly told about it
in that manner.

--

      OS/2 + BeOS + Linux + Solaris + Win95 + WinNT4 + FreeBSD + DOS
       + VMWare + Fusion + vMac + Executor = PC Hobbyist Heaven! :-)
        Real programmers use ER SNAP$ and read Fieldata (in octal!)

Quote:> There is a minor problem in that with some BIOSes Linux does not correct
> discover memory above 64Mb, so it only uses the first 64.  If you have
> that problem you just have to add an append argument to your lilo.conf
> file to tell the kernel explicitly how much memory it has and all is
> fine.

Yes - that I knew about already, and overcame.  I just wondered if there
was a loading issue.  I'd heard rumbled somewhere that there was.
--
Raj Rijhwani        (umtsb5/16) |  This is the voice of the Mysterons...

                                |  "Lieutenant Green:  Launch all Angels!"
http://www.courtfld.demon.co.uk/raj/ (demon, and gods, willing...)

On Tue, 09 May 2000 20:48:58 GMT,

There is on some Intel chipsets, which refuse to cache more than the
first 64M of RAM.  Things that load in the 'uncachable' memory would be
slower.  To my knowledge, this only affects chipsets for the plain old
Pentium (ie, socket 7) line, so if you're using a Pentium Pro, it
shouldn't be a problem.

--
Brian Moore                       | Of course vi is God's editor.
      Sysadmin, C/Perl Hacker     | If He used Emacs, He'd still be waiting
      Usenet Vandal               |  for it to load on the seventh day.
      Netscum, Bane of Elves.

IIRC SETI relies heavily on the FPU so I'd guess the K6-2 has a
weak FPU.

Quote:>What am I missing?  What do people recommend in the way of "proper"
>benchmarks?  Can I log kernel stats and perform the equivalent of
>sar reports on commercial platforms?

Dunno about "proper", but I find the speed test from the OpenSSL
package to be useful for gauging integer performance.

I happen to have upgraded a Cyrix MII 233MHz to an AMD K6/2 500MHz
and it's about 2.5 times faster.

Comparison with a PII indicates integer performance akin to a 333MHz
device.

Quote:>(I heard some rumblings about Linux not behaving well with more than
>64Mb of RAM.  Since the new machine has 128Mb I wonder, is this a true
>concern?)

VX and TX Pentium support chipsets from Intel could only cache
the first 64M of memory. Dunno if that's what your thinking of.

Jon Plews.

Quote:> IIRC SETI relies heavily on the FPU so I'd guess the K6-2 has a
> weak FPU.

This would certainly seem to be the case.  (Now regretting having
bought AMD - something I would never have thought I would say...)

Quote:> VX and TX Pentium support chipsets from Intel could only cache
> the first 64M of memory. Dunno if that's what your thinking of.

Given the implications, I'm guessing it is (and thanks for the answers
peeps).
--
Raj Rijhwani        (umtsb5/16) |  This is the voice of the Mysterons...

                                |  "Lieutenant Green:  Launch all Angels!"
http://www.courtfld.demon.co.uk/raj/ (demon, and gods, willing...)

This is something I thought I heard once or twice before - AMD K6-* FPUs
might not be quite up to scratch compared to the Intel ones, or something
like that. In any case, it depends what you want to do - I know that for
having a notebook sitting onna knee recompiling a kernel, I'd choose AMD
any day. (Or rather, a G3, but so it goes ;)

~Tim
--
| Geek Code: GCS dpu s-:+ a-- C++++ UBLUAVHSC++++ P+++ L++ E--- W+++(--) N++
| w--- O- M-- V-- PS PGP++ t--- X+(-) b D+ G e++(*) h++(*) r--- y-          
| The sun is melting over the hills,         | http://piglet.is.dreaming.org/

Yes, the k6-2s don't have very good FP units on them but they do perform
very admirably at integer (probably as good as if not better than a P2).

The k6-3s made up for it a lot but they weren't really worth it with athlon
just around the corner, though I'm still kicking myself that I didn't get a
k6-3-400 instead of the k6-2-400 I've got in this here laptop.

PJF

Also, setiathome has been pretty heavily optimised for Intel CPUs; the
K6/2's FPU isn't so hot, but it's by no means that slow.
--

It moves between us, for one moment, like * in your heart, with remedies
from the ancient gods, to heal the morals of our shadows. Devil, come to me,
open up the door, lead me ciahra to the centre of it all...(FotN:Submission)