Monitoring server for crashes

[Ian Smith]

In freebsd-questions Digest, Vol 636, Issue 7, Message: 10
On Fri, 12 Aug 2016 11:51:50 -0400 Robert Fitzpatrick <robert at webtent.org> wrote:
 > Valeri Galtsev wrote:
 > > Before doing such monitoring I would really do a good hardware test.
 > > Incidentally, who is hardware manufacturer (just for my curiosity). The
 > > usual suspects are: memory (poor/flaky memory, or combination of memory
 > > with slightly different specs; these even though they may work together
 > > can lead to failure sometimes very rarely, like once every 6 Months which
 > > is really hard to troubleshoot: just avoid this). Another possibility:
 > > tripping temperature threshold set in BIOS. (These, BTW will leave no
 > > tracks in crash, logs etc.) Check this and bring threshold some 15-20 F (7
 > > - 10 C ) up.  Incidentally: which CPU(s) do you have? (I'm used to think,
 > > AMD will withstand any abuse without failing: you almost can boil water on
 > > these, Intels are not as robust). What I would do is : open the box, leave
 > > minimal hardware (run with minimal amount of RAM, remove all extra cards
 > > etc) and see if you have problem with this minimal hardware configuration.
 > > If not, start adding hardware, install all RAM first, test if it doesn't
 > > crash. Run memtest96 at this point for at least 48 hours (or at the very
 > > minimum 2-3 full loops of test). In this configuration try to run system
 > > and create significant CPU load (several multi-thread "build world" can
 > > help do that), and simultaneously try to use all the RAM. Things are
 > > slightly different under heavy load. And so on - add the rest of hardware
 > > and test... One more thing: check if your PS provides at least 30% more
 > > power than all hardware may need. Marginally insufficient power may lead
 > > to unpredictable thing on PCI bus. Incidentally, how old is power supply
 > > (and the rest of hardware). Electrolytic capacitors may loose capacitance
 > > with age, thus not filtering well enough ripple on PS leads (capacitors
 > > inside PS), on CPU power leads and on PCI bus power lines (capacitors on
 > > system board - check if they do not showing traces of leakage).

All good advice Valeri; not sure about messing with temps in BIOS though 
.. FreeBSD should be handling that ok via ACPI thermal Zones (versus 
_HOT and _CRT temperatures) which should cleanly shutdown at _CRT temp.
That said, if it gets anywhere near that hot there's a serious issue ..

 > Thanks for all the suggestions, will check temp and other info in BIOS 
 > tonight, I really can't have the server down for long memory test, will 
 > make sure all memory is the same. The server is IBM x3650 with 2 Quad 
 > Core Xeon L5420 a mixture of drives using hardware ServeRAID 8k and 12GB 
 > of RAM. I purchased second hand in 2011. I have a screenshot of the 
 > product data screen in the BIOS, it has a diagnostics date of Aug 2009 
 > in the BIOS, all hardware should be original except drives and memory. 
 > The load comes from a PostgreSQL database primarily, also provides DNS 
 > and LDAP services. Not sure heat is the issue, mainly happens at the 
 > same general time at night, heaviest load is definitely during the day.

I guess you've checked with ibm re a BIOS update .. 2009 is a while ago.

Apart from RAM, which rarely just 'goes bad' esp. on server grade gear, 
but "rarely happens" happens too.

First thing I'd suspect at that age would be the power supply - can you 
swap it with another?  Quickest fix if it works - and it was needed.

Second would be temperature, possibly fan/s - which is also the primary 
cause of blown P/S in my experience.  Below is a script I run from cron 
from 02:59 through 3:09 to record load averages and temperatures through 
daily maintenance from 3:01, every 10 seconds - for debugging a load 
average issue, not relevant here.  Or you can run it over SSH at home, 
and read the last entries over breakfast, whether it crashes or not ..

The lack of any messages - and you should see one if ACPI thermal zone 
detection and forced shutdown is working properly - suggests more of a 
hardware seizure, but at 10 second testing you could see if temps (and 
load) were a problem prior to crash, at least if it happens in a window.

 > I see now, most of the time it happens during dumping of the db each 
 > night, but it has happened once during the day and once a couple of 
 > hours before backup. I'm leaning toward a memory issue and will 
 > definitely visit the data center tonight and see the types. The db size 
 > has not changed much over time and this just started recently. It is a 
 > SpamAssassin/ClamAV db and purges, vacuums every night after dumping. I 
 > will disable and do dump manually tonight, 90% of the time it seems to 
 > be going down during backup of the largest db. Perhaps the crashes have 
 > caused a table to corrupt, I 'fsck -y' all mounts in single user mode 
 > after every crash.

Do the fscks log success or any problems then?  If not, might be worth 
doing manual fsck to check?

/etc/crontab:
59      2       *       *       *       root    /root/bin/loadavg_daily

/root/bin/loadavg_daily:
=======
#!/bin/sh
# 19Feb16 loadavg_daily .. every 10 seconds from 02:59 to 03:09 (run by cron)
log='/root/loadavg_daily.log'
secs=10
i=0
/root/bin/x200stat >> $log	# or something else, or nothing ..
while [ $i -lt 60 ]; do
        echo -n "`uptime`  " >> $log
        echo "`sysctl -n hw.acpi.thermal.tz0.temperature`" \
        "`sysctl -n hw.acpi.thermal.tz1.temperature`" >> $log
        sleep $secs
        i=$((i + 1))
done
/root/bin/x200stat >> $log
echo >> $log
=======

Check sysctl hw.acpi.thermal for your thermal zones of interest.

HTH, Ian