Need help with lockups on tyan s2460 motherboard in SMP mode

Need help with lockups on tyan s2460 motherboard in SMP mode

[James Strandboge]

I have googled a lot on this issue, but if there is somewhere else I can
look, please tell me.

I have a dual AMD 1600+ (MP processors) system with tyan s2460
motherboard that freezes with no error messages in syslog.  I have read
(among a lot of other things):
http://ouray.cudenver.edu/~etumenba/smp-howto/SMP-HOWTO.html

and what I have done so far is:

replaced power supply with Enermax 550
BIOS upgrade to 1.05
kernels used are 2.4.18 from debian, 2.4.20 and 2.4.21-pre6 
compiled in acpi (acpi-200303028-2.4.21-pre6)
compiled out acpi
compiled with and without Athlon support
Currently using 2.4.21-pre6 for i386

booted with combinations of the following:
noapic
noacpi
acpi=off
nopentium

tried the following bios settings in various combinations:
MultiProcessor specification 1.1 and 1.4
Use PCI Interrupt Entries in MP as yes and no
enabled/disabled acpi
enabled/disabled power management

The only way I have found to reliably crash the system is to compile two
kernels simultaneously while in smp mode.  If I boot with 'nosmp' the
system seems ok (still running and compiling after 9 hours).

Because the system runs ok with one processor, it leads me to think it
may be kernel code, and not heat or memory related.  Though I guess it
could be the motherboard still.  Is this a correct assessment?  I'd be
happy to send any additional information.

I am new to SMP, so any and all suggestions are welcome.

Thanks,

Jamie Strandboge

-- 
James Strandboge
Targeted Performance Partners, LLC
Web: http://www.tpptraining.com
E-mail: jamie@tpptraining.com
Tel: (585) 271-8370
Fax: (585) 271-8373

Re: Need help with lockups on tyan s2460 motherboard in SMP mode

[James Strandboge]

On Tue, 2003-04-22 at 09:38, James Strandboge wrote:

> The only way I have found to reliably crash the system is to compile two
> kernels simultaneously while in smp mode.  If I boot with 'nosmp' the
> system seems ok (still running and compiling after 9 hours).

Well, it crashed after 15 hours with nosmp.  Guess its hardware related.

Jamie
-- 
James Strandboge
Targeted Performance Partners, LLC
Web: http://www.tpptraining.com
E-mail: jamie@tpptraining.com
Tel: (585) 271-8370
Fax: (585) 271-8373

Re: Need help with lockups on tyan s2460 motherboard in SMP mode

[James Strandboge]

On Tue, 2003-04-22 at 09:38, James Strandboge wrote:
> I have a dual AMD 1600+ (MP processors) system with tyan s2460
> motherboard that freezes with no error messages in syslog.  

First off, thank you to everyone who responded-- your input helped me
figure this out.  I wanted to send this to the lists so others might
benefit.

As it turns out, the problem appears to have been solely heat related. 
I ended up installing a vanilla 2.4.20 kernel with smp for 386 (ie no
pentium or athlon optimizations), without acpi, but with apm as a
module.  I disabled power management in the BIOS, and boot with only
apm=power-off for kernel command line arguments.  Notice I did NOT use
noapic.  The kernel boots fine and has no APIC errors (though there is
that errata #22 message).  I then installed lm_sensors 2.7.0 and
i2c-2.7.0 so I could read the temperature.

The temperature readings for the 3904 transistor are the ones to use,
since they are closest to those in the BIOS.  I called tyan and spoke
with someone on the phone, and he said that the system will lock up
around 60C (not the advertised 80-90C, because the sensor is reading air
temperature).  With the case all closed up I ran two kernel compiles,
and two 'top -d .01' instances all at once.  Then I ran 'sensors' every
30 seconds, and sure enough, the computer got near 60C and crashed.  I
took the sides off the case, and used an oscillating fan to blow air
into the computer and rebooted.  'sensors' now shows the temperature
between 39-43C and has been running the above stress test for several
hours (it has never run it for that long).

So now I will be buying some additional fans and making sure the air
flow is good in the case.  I also whole heartedly recommend using
lm_sensors.  It REALLY helped with debugging this issue.

Jamie

-- 
James Strandboge
Targeted Performance Partners, LLC
Web: http://www.tpptraining.com
E-mail: jamie@tpptraining.com
Tel: (585) 271-8370
Fax: (585) 271-8373