Very Computer

I just built a new box with:

-  Tyan Tomcat 3D with 512K cache
-  2 Pentium 100, an old one with stepping 5 and a new one with
   stepping 6
-  32M EDO memory
-  53c825 controller with 2G wide SCSI drive.

Building the 2.1.6 kernel takes 12 minutes regardless of the number of
CPUs.  I was expecting some speedup with dual CPUs, but found none.

The motherboard manual says, "When using two Pentiums, Intel
recommends that the chips have the same s-spec number..."  What's
this "s-spec"?  Is it the same as "stepping"?  What is the stepping
number anyway?  Do I really have to have the same stepping for both
CPUs?

I understand that SMP is built into the 2.X.X kernels, that is,
there's no SMP configuration option.  (I also tried Windows NT 4.0
with my new box.  It didn't see the second CPU at all.  I recall
there's a patch that must be applied separately from the regular
installation.)

Please help.  Any comments will be appreciated.  Here are the contents
of /proc/cpuinfo in 3 different cases:

1.  One CPU
processor       : 0
cpu             : 586
model           : Pentium 75+
vendor_id       : GenuineIntel
stepping        : 5
fdiv_bug        : no
hlt_bug         : no
fpu             : yes
fpu_exception   : yes
cpuid           : yes
wp              : yes
flags           : fpu vme de pse tsc msr mce cx8
bogomips        : 39.73

2.  Two CPUs
processor       : 0
cpu             : 586
model           : Pentium 75+
vendor_id       : GenuineIntel
stepping        : 5
fdiv_bug        : no
hlt_bug         : no
fpu             : yes
fpu_exception   : yes
cpuid           : yes
wp              : yes
flags           : fpu vme de pse tsc msr mce cx8 apic
bogomips        : 39.73

3.  Two CPUs, slots reversed
processor       : 0
cpu             : 586
model           : Pentium 75+
vendor_id       : GenuineIntel
stepping        : 6
fdiv_bug        : no
hlt_bug         : no
fpu             : yes
fpu_exception   : yes
cpuid           : yes
wp              : yes
flags           : fpu vme de pse tsc msr mce cx8 apic
bogomips        : 39.73

What's the "apic" flag?  The Tyan motherboard manual uses the same
acronym without explaining it.

--
Lam Dang

: -  2 Pentium 100, an old one with stepping 5 and a new one with
:    stepping 6
: Building the 2.1.6 kernel takes 12 minutes regardless of the number of
: CPUs.  I was expecting some speedup with dual CPUs, but found none.
with two p133 it takes you ~7 min for recent Kernels. You have to enable
SMP-support by uncommenting :
#SMP = 1
in /usr/src/linux/Makefile

: The motherboard manual says, "When using two Pentiums, Intel
: recommends that the chips have the same s-spec number..."  What's
: this "s-spec"?  Is it the same as "stepping"?  What is the stepping
: number anyway?  Do I really have to have the same stepping for both
: CPUs?
dont know, i use SSS Pentiums whith the same stepping. You drive it with one
CPU. Cat /proc/cpuinfo has to report two cpus and you get ~105bogomips with
2xp133.

W.
--

  Ein UART 16550A ist ein Muss fuer DFUE. SCSI ist der Koenigsweg.

: : Building the 2.1.6 kernel takes 12 minutes regardless of the number of
                 ^^^^^
mumpf just seen it. 2.1.6 wont work with SMP, it hangs at boot. All 2.1.x
prior 2.1.16 wont work with SMP enabled. Get the latest patches, or use
2.0.x.

W.

--

  Ein UART 16550A ist ein Muss fuer DFUE. SCSI ist der Koenigsweg.

Following the advice in your earlier post, I uncommented SMP=1 in
/usr/src/linux/Makefile.  Now the kernel recognizes both CPUs as seen
in the boot messages and /proc/cpuinfo.  Thanks.

But I still don't see a speed gain in building 2.1.16.  I used "make -j
4" in both cases; I didn't see a lot of disk activity (light and sound)
in either case.  Any other suggestions?

--
Lam Dang

Quote:

> Just calling make with a '-j 4' won't work, because the top level
> kernel Makefile defines make as 'make' and the '-j 4' never gets
> passed down to lower level Makefiles.

> Edit your top level Makefile and change:

> MAKE       =make

> to

> MAKE       =make -j 4

> and then you should see the speed gain. ;)

I also tried "make -j" without a number as Wofgang suggested.  Now the
disk wouldn't stop, probably because I've got only 32M of memory.
Then I tried "make -j -l 0.75" to cut down on the number of jobs based
on load.  The disk didn't thrash, but there was no speedup.

I still want to know what "s-spec" and "stepping" numbers are.
Obviously different stepping numbers don't seem to matter for SMP.

--
Lam Dang

Lam> Sorry it's a typo.  I meant 2.1.16.

Lam> Following the advice in your earlier post, I uncommented SMP=1 in
Lam> /usr/src/linux/Makefile.  Now the kernel recognizes both CPUs as
Lam> seen in the boot messages and /proc/cpuinfo.  Thanks.

Lam> But I still don't see a speed gain in building 2.1.16.  I used
Lam> "make -j 4" in both cases; I didn't see a lot of disk activity
Lam> (light and sound) in either case.  Any other suggestions?

Just calling make with a '-j 4' won't work, because the top level
kernel Makefile defines make as 'make' and the '-j 4' never gets
passed down to lower level Makefiles.

Edit your top level Makefile and change:

MAKE    =make

to

MAKE    =make -j 4

and then you should see the speed gain. ;)

kevin
--


WWW -> http://scrye.com/~kevin/

: But I still don't see a speed gain in building 2.1.16.  I used "make -j
: 4" in both cases; I didn't see a lot of disk activity (light and sound)
: in either case.  Any other suggestions?
Hm maybe make is the Problem mine ignores digits given to -j in bash.
Try "time make -j zImage". You should see lots of gccs in memory and much
disk activity. Hope there is enough swap-space. Tough guys edit the top-level
Makefile:
Make = make -j
and type make -j zImage so all jobs run simultaneously even in sub-
directories.
Much memory needed ~120MB so your UW hostadapter can play at his best..

W.
--

  Ein UART 16550A ist ein Muss fuer DFUE. SCSI ist der Koenigsweg.

Quote:

> I understand that SMP is built into the 2.X.X kernels, that is,
> there's no SMP configuration option.  (I also tried Windows NT 4.0
> with my new box.  It didn't see the second CPU at all.  I recall
> there's a patch that must be applied separately from the regular
> installation.)

| But I still don't see a speed gain in building 2.1.16.  I used "make -j
| 4" in both cases; I didn't see a lot of disk activity (light and sound)
| in either case.  Any other suggestions?

make zImage modules MAKE='make -j 3'

-j3 gives me 198% cpu with dual PPro-200 and 64M. I've heard people
speak of larger values, but trying them didn't help measurably.

Even with all that a kernel and module build can take 3-6 minutes
from make clean. It depends on how many things you put in your
kernel ;-)
--

  What do you mean I shouldn't do thing like that at my age?
  At my age if you don't do things like that you might die of natural causes!

Quote:> When you install NT, it automaticaly installs the smp support if it
> finds 2 CPUs.

I'm quite happy with the whole thing now.  I've got both Linux 2.X.X
and NT 4.0 running SMP.  The NT Task Manager can create load and graph
how each of the CPUs responds to varying load at the thread
level.  It's fun playing around with this SMP stuff.  And I didn't
have to spend a lot for it either.

Thank you all for your help.

--
Lam Dang

Quote:>You're right.  When I reinstalled NT 4.0 with both CPUs present, I got
>SMP.  If I installed SP2 _and_ networking (TCP/IP and RAS), however,
>it would crash soon after the NT logo showed up.  With only SP1, it's
>working all right.

M$ needs to get their act together.  Releasing a SP that crashes their
"mission critical" OS isn't going to please many admins or users.

P.S.  I spent approx $700 for NT Svr V.4 (for home use when I'm not
using my favorite OS - Linux).  Contacted M$ for "complementary"
installation support for this problem.  They wouldn't talk to me
unless I gave them my Visa for a $190 support charge.  And Bill Gates
expects to have NT on every desktop?