Hi Martin,

Actually the system check is just after the web server setup and before the
apache checks, I have basically modified the default monitrc file and have
not changed the order of the checks. So my check order is as follows.

check system
check apache_bin
check httpd
check postfix
check other services

if there is a change to the config i can make to remedy this i will be
happy to try it and report the results.

Regards

Wayne



On 8 December 2011 13:50, Martin Pala <[email protected]> wrote:

>  Thanks, the root cause is clear now.
>
> It seems that your configuration (most probably the apache check) uses the
> pattern based process check and the system check is most probably defined
> behinf the apache in monitrc. When you do restart of such service, monit
> waits for it to start and refreshes the process list to see whether it
> started. The process list refresh also refreshes the cpu usage statistics -
> this happens every 5 milliseconds until the process starts or the action
> times out.
>
> The CPU usage reported by monit (for example system 50%) is thus true but
> the value comes from very short timeframe (cpu usage from last 5
> milliseconds) instead of full cycle (for example cpu usage from last 5
> seconds) . If the system check will be defined first (in front of apache
> check), this won't happen, as it will take the initial values (from the
> cycle start) before the apache action occurred.
>
> => it is bug limited to specific configuration:
> 1.) "check process myproc matching …" is used
> 2.) "check system" is defined after the myproc check
> 3.) the myproc service is restarted
>
> Workaround:
> move the "check system" ahead of other services in your monit
> configuration file
>
>
> We'll fix the problem.
>
> Thanks for help with the testing and data :)
>
> Regards,
> Martin
>
>
>
>  On Dec 8, 2011, at 2:03 PM, Lawrence, Wayne wrote:
>
>  Hi Martin,
>
> did as you instructed here is the output.
> From my untrained eye there is some serious miscalculation in the 4th
> CPUDEBUG statement not a clue how it arrives at that figure.
>
> CPUDEBUG: used_system_memory_sysdep: time=1323349102: cpu_user=293199
> (-1.00%), cpu_nice=547, cpu_syst=194433 (-1.00%), cpu_idle=58991209,
> cpu_wait=31605 (-1.00%), cpu_irq=153, cpu_softirq=1990, cpu_total=59512589
> -- old_cpu_user=0, old_cpu_syst=0, old_cpu_wait=0, old_cpu_total=0
> CPUDEBUG: check_system: time=1323349102:
> systeminfo.total_cpu_user_percent=-1.00%,
> systeminfo.total_cpu_syst_percent=-1.00%,
> systeminfo.total_cpu_wait_percent=-1.00%
> CPUDEBUG: used_system_memory_sysdep: time=1323349142: cpu_user=293227
> (0.70%), cpu_nice=547, cpu_syst=194469 (0.90%), cpu_idle=58995131,
> cpu_wait=31606 (0.00%), cpu_irq=153, cpu_softirq=1990, cpu_total=59516576
> -- old_cpu_user=293199, old_cpu_syst=194433, old_cpu_wait=31605,
> old_cpu_total=59512589
> CPUDEBUG: used_system_memory_sysdep: time=1323349142: cpu_user=293227
> (-214748364.80%), cpu_nice=547, cpu_syst=194469 (-214748364.80%),
> cpu_idle=58995131, cpu_wait=31606 (-214748364.80%), cpu_irq=153,
> cpu_softirq=1990, cpu_total=59516576 -- old_cpu_user=293227,
> old_cpu_syst=194469, old_cpu_wait=31606, old_cpu_total=59516576
> CPUDEBUG: used_system_memory_sysdep: time=1323349142: cpu_user=293229
> (0.00%), cpu_nice=547, cpu_syst=194473 (100.00%), cpu_idle=58995132,
> cpu_wait=31606 (0.00%), cpu_irq=153, cpu_softirq=1990, cpu_total=59516583
> -- old_cpu_user=293229, old_cpu_syst=194472, old_cpu_wait=31606,
> old_cpu_total=59516582
> CPUDEBUG: check_system: time=1323349142:
> systeminfo.total_cpu_user_percent=0.00%,
> systeminfo.total_cpu_syst_percent=100.00%,
> systeminfo.total_cpu_wait_percent=0.00%
> CPUDEBUG: used_system_memory_sysdep: time=1323349202: cpu_user=293307
> (0.90%), cpu_nice=547, cpu_syst=194542 (0.90%), cpu_idle=59001021,
> cpu_wait=31610 (0.00%), cpu_irq=153, cpu_softirq=1990, cpu_total=59522623
> -- old_cpu_user=293252, old_cpu_syst=194483, old_cpu_wait=31606,
> old_cpu_total=59516616
> CPUDEBUG: check_system: time=1323349202:
> systeminfo.total_cpu_user_percent=0.90%,
> systeminfo.total_cpu_syst_percent=0.90%,
> systeminfo.total_cpu_wait_percent=0.00%
>
> Regards
>
> Wayne
>
> On 8 December 2011 12:50, Martin Pala <[email protected]> wrote:
>
>>  Hi,
>>
>> thanks for update. I have prepared the debug version, which logs the
>> values computed based on /proc/stat right when they are ready and once
>> again before the values are checked, so we can see whether the values were
>> read+computed correctly and whether no memory corruption occurred before
>> they were compared by the validation engine => there are two "CPUDEBUG" log
>> entries per cycle.
>>
>> You can get it here: http://www.mmonit.com/tmp/monit-5.3.1p2.tar.gz
>>
>> To compile:
>> tar -xzf monit-5.3.1p2.tar.gz
>> cd monit-5.3.1p2
>> ./configure
>> make
>>
>> Then stop existing monit instance and run new monit binary:
>> ./monit -vI  2>&1 | grep CPUDEBUG
>>
>> after you'll replicate the problem, terminate monit with ^C and send the
>> whole CPUDEBUG output since monit start
>>
>> Regards,,
>> Martin
>>
>>
>>
>>  On Dec 8, 2011, at 11:39 AM, Lawrence, Wayne wrote:
>>
>>  Hi Martin just as a side note here i disabled the cpu ssystem test and
>> tried again and it seems that the issue is present with all the cpu
>> monitoring/
>>
>> I used the restarting of httpd as i knew it would trigger and alert and
>> these were the results.
>>
>>  Date:        Thu, 08 Dec 2011 10:27:59
>>       Action:      alert
>>       Host:        <hostname removed>
>>       Description: cpu user usage of 100.0% matches resource limit [cpu
>> user usage>70.0%]
>>
>>
>> I ran vmstat 1 10 at the same time as you can see its the 4th line.
>>
>>
>> procs -----------memory---------- ---swap-- -----io---- --system--
>> -----cpu-----
>>  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy
>> id wa st
>>  0  0      0 739220 142536 973532    0    0     4     7   10    6  0  0
>> 99  0  0
>>  0  0      0 739088 142536 973532    0    0     0     0  114  160  0  1
>> 99  0  0
>>  3  0      0 739088 142536 973536    0    0     0     0  126  169  1  2
>> 97  0  0
>>  0  0      0 737336 142536 973544    0    0     0   168  721  796 35 14
>> 50  1  0
>>  1  0      0 736964 142536 973544    0    0     0     0  109  160  1  1
>> 98  0  0
>>
>> and just to make it a little simpler i ran sar 1 10 as well as it is more
>> human readable.
>>
>> 10:27:55        CPU     %user     %nice   %system   %iowait    %steal
>> %idle
>> 10:27:56        all      1.01      0.00      1.01      0.00      0.00
>> 97.98
>> 10:27:57        all      0.00      0.00      1.00      0.00      0.00
>> 99.00
>> 10:27:58        all      3.96      0.00      3.96      0.00      0.00
>> 92.08
>> 10:27:59        all     32.00      0.00     12.00      1.00      0.00
>> 55.00
>>
>> Something struck me as odd while testing this yesterdays results
>> reporting 50% system usage from 15.84% actual means the reported usage is
>> 3.2 times the actual. todays reported user usage of 100% is 3.2 times the
>> actual 32%. so it seems just need to work out why it is multiplying the
>> results.
>>
>> Regards
>>
>> Wayne
>>
>> On 7 December 2011 11:43, Lawrence, Wayne <[email protected]>wrote:
>>
>>> Hi Martin,
>>>
>>> I downloaded the source from the Monit website and compiled it on the
>>> server.
>>> I have started monit in verbose mode and this is the
>>> relevant information it outputs when the event occurs.
>>>
>>>   cpu system usage of 50.0% matches resource limit [cpu system
>>> usage>30.0%]
>>>
>>> -------------------------------------------------------------------------------
>>>    ../tools/bin/monit() [0x41a533]
>>>  ../tools/bin/monit(LogError+0x9f) [0x41ad2f]
>>>    ../tools/bin/monit(Event_post+0x328) [0x417ba8]
>>>     ..t/tools/bin/monit() [0x428071]
>>>     ../tools/bin/monit(check_system+0x2b) [0x4285bb]
>>>     ../tools/bin/monit(validate+0x226) [0x42ad16]
>>>    ../tools/bin/monit() [0x41422d]
>>>     ../tools/bin/monit(main+0x511) [0x4149e1]
>>>     /lib64/libc.so.6(__libc_start_main+0xfd) [0x3592c1ecdd]
>>>     ../tools/bin/monit() [0x40b179]
>>>
>>> -------------------------------------------------------------------------------
>>> Unfortunately remote access is not an option but I will happily run a
>>> debug version to try and track down this problem as I really would like to
>>> use Monit for my current build.
>>>
>>> Regards
>>>
>>> Wayne
>>>  On 7 December 2011 11:17, Martin Pala <[email protected]> wrote:
>>>
>>>>   Thanks for data.
>>>>
>>>> The /proc/stat format is this:
>>>>
>>>>     cpu <user> <nice> <system> <idle> <wait> <irq> <softirq>
>>>>
>>>> The values count the cpu cycles, so if we subtract the corresponding
>>>> values from your output, we get this:
>>>>
>>>>                    user   nice   system   idle   wait   irq   softirq
>>>> |   total
>>>> 09:57:35    1         0        1              99     0       0      0
>>>>        |    101
>>>> 09:57:36    1         0        0              98     0       0      0
>>>>        |    99
>>>> 09:57:37    25       0        16           59     1       0      0
>>>>      |    101
>>>>  09:57:38    1         0        2              98     0       0
>>>> 0          |    101
>>>>
>>>> => at  09:57:37 the cpu usage was:
>>>>
>>>> user      = 24.75%
>>>> system =  15.84%
>>>> wait      =   0.99%
>>>>
>>>> This corresponds to the previous vmstat output. Monit counts the cpu
>>>> usage the same way as above and doesn't modify these values => your monit
>>>> really reports strange cpu usage (reported 50% vs. real ~ 16%).
>>>>
>>>> What's the origin of your monit binary? Did you compile it from
>>>> original source code or some 3rd party source code distibution? (such as
>>>> RHEL or Fedora repository). Or do you use the pre-compiled binaries from
>>>> www.mmonit.com? Or some 3rd party binary, patches or source code from
>>>> other site?
>>>>
>>>> Please can you try to run monit in verbose mode and provide full
>>>> output?:
>>>>
>>>>    1.) stop monit
>>>>    2.) run monit in foreground with verbose mode enabled:
>>>>        ./monit -vI
>>>>    3.) after the problem happens, stop monit with "^C" and send output
>>>>
>>>> I can also prepare debug version which will dump the cpu usage related
>>>> informations or if you can provide remote access to the system, i can
>>>> troubleshoot the problem remotely.
>>>>
>>>>
>>>> Regards,
>>>> Martin
>>>>
>>>>
>>>>
>>>>   On Dec 7, 2011, at 11:07 AM, Lawrence, Wayne wrote:
>>>>
>>>>  Hi Martin,
>>>>
>>>> this is the output of the commands you requested.
>>>>
>>>> 1.) uname -m
>>>>
>>>> x86_64
>>>>
>>>>  2.) file `which monit`
>>>>
>>>>  ELF 64-bit LSB executable, x86-64, version 1 (GNU/Linux), dynamically
>>>> linked (uses shared libs), for GNU/Linux 2.6.18, not stripped
>>>> I ran the command you supplied to get the cup usage directly as well
>>>> while restarting the httpd service as i know this will generate an alert.
>>>>
>>>>
>>>>
>>>>       Date:        Wed, 07 Dec 2011 09:57:37
>>>>       Action:      exec
>>>>       Host:        <hostname removed>
>>>>       Description: cpu system usage of 50.0% matches resource limit
>>>> [cpu system usage>30.0%]
>>>>
>>>> Wed Dec  7 09:57:34 GMT 2011
>>>> cpu  207060 501 103542 49452254 25303 83 1569 0 0
>>>> Wed Dec  7 09:57:35 GMT 2011
>>>> cpu  207061 501 103543 49452353 25303 83 1569 0 0
>>>> Wed Dec  7 09:57:36 GMT 2011
>>>> cpu  207062 501 103543 49452451 25303 83 1569 0 0
>>>> Wed Dec  7 09:57:37 GMT 2011
>>>> cpu  207087 501 103559 49452510 25304 83 1569 0 0
>>>> Wed Dec  7 09:57:38 GMT 2011
>>>> cpu  207088 501 103561 49452608 25304 83 1569 0 0
>>>> Wed Dec  7 09:57:40 GMT 2011
>>>> If my understanding of /proc/stat is coreect this still doesnt make any
>>>> sense but i may be wrong.
>>>>
>>>> Regards
>>>>
>>>> Wayne
>>>>
>>>>
>>>>
>>>> On 7 December 2011 09:37, Martin Pala <[email protected]> wrote:
>>>>
>>>>>  Please can you check that your monit binary matches the system
>>>>> architecture? (i.e. for example 64-bit monit binary on 64-bit system - not
>>>>> 32-bit monit on 64-bit system)
>>>>>
>>>>> To verify provide please the output of following commands:
>>>>> 1.) uname -m
>>>>> 2.) file `which monit`
>>>>>
>>>>> Monit takes the statistics from the /proc/stat kernel interface. You
>>>>> can collect the statistics manually like this - for example to fetch the
>>>>> state in 1 second intervals (30 samples):
>>>>>
>>>>>  $ for ((i=0; i<30; i++)); do date; grep "cpu " /proc/stat; sleep 1;
>>>>> done
>>>>>
>>>>> Note: monit takes the first /proc/stat line ("cpu") which contains the
>>>>> overall cpu usage in the system (summary of all cpus). The /proc/stat also
>>>>> contains per-cpu statistics if you want to collect all the statistics,
>>>>> replace the "grep 'cpu '" simply with "cat".
>>>>>
>>>>> Regards,
>>>>> Martin
>>>>>
>>>>>
>>>>>  On Dec 7, 2011, at 10:04 AM, Lawrence, Wayne wrote:
>>>>>
>>>>>  Hi Martin,
>>>>>
>>>>> I have tried various methods to dientify the cause of this and took
>>>>> your advice and used vmstat. I simply restarted the httpd process from the
>>>>> monit web interface while the comand was running and got the following
>>>>> warning.
>>>>>
>>>>>        Description: cpu system usage of 50.0% matches resource limit
>>>>> [cpu system usage>30.0%]
>>>>>
>>>>> But vmstat doesnt show that level of usage at the point of alert. As
>>>>> you can see there is some usage in the 3rd line of the output when i
>>>>> restarted the httpd service but it doesnt seem enough to trigger an alert.
>>>>>
>>>>> vmstat 1 10
>>>>> procs -----------memory---------- ---swap-- -----io---- --system--
>>>>> -----cpu-----
>>>>>  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us
>>>>> sy id wa st
>>>>>  0  0      0 859596 114684 856908    0    0     4     6   81   77  0
>>>>> 0 99  0  0
>>>>>  0  0      0 859448 114684 856916    0    0     0     0  100   94  1
>>>>> 0 99  0  0
>>>>>  0  0      0 898352 114692 815600    0    0     0   168  555  605 23
>>>>> 15 61  1  0
>>>>>
>>>>> Not sure if there are any other tests i can run to narrow this down a
>>>>> bit further as it still isn't making sense.
>>>>>
>>>>> Regards
>>>>>
>>>>> Wayne
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 7 December 2011 08:27, Martin Pala <[email protected]> wrote:
>>>>>
>>>>>>  Hi Lawrence,
>>>>>>
>>>>>> the test which triggers the alert is "system" cpu => it's the time
>>>>>> the system spend in kernel mode. The cpu usage could be triggered by some
>>>>>> background kernel task, to verify the monit report matches the system cpu
>>>>>> usage, you should use either "vmstat" or "top" instead of "ps".
>>>>>>
>>>>>> Best regards,
>>>>>> Martin
>>>>>>
>>>>>>
>>>>>>  On Dec 6, 2011, at 1:19 PM, Lawrence, Wayne wrote:
>>>>>>
>>>>>>  Hi Igor,
>>>>>>
>>>>>> the operating system is RHEL6 and monit version is 5.3.1
>>>>>>
>>>>>> this is what i have in my config
>>>>>>
>>>>>>     if cpu usage (user) > 70% then alert
>>>>>>     if cpu usage (system) > 30% then alert
>>>>>>     if cpu usage (wait) > 20% then alert
>>>>>>
>>>>>> this is one of the errors
>>>>>> Description: cpu system usage of 50.0% matches resource limit [cpu
>>>>>> system usage>30.0%]
>>>>>>
>>>>>> this is what i get in /var/log/messages
>>>>>> Dec  6 12:01:29 <hostname-removed> monit[864]: <hostname-removed> cpu
>>>>>> system usage of 50.0% matches resource limit [cpu system usage>30.0%]
>>>>>> Dec  6 12:02:29 <hostname-removed> monit[864]:
>>>>>> <hostname-removed><hostname-removed>' cpu system usage check succeeded
>>>>>> [current cpu system usage=0.9%]
>>>>>>
>>>>>> this is the output of ps --no-headers -A -o "%*cpu* sz ucomm" | sort
>>>>>> -k1nr | head -20
>>>>>>
>>>>>>  12:01:29 up 4 days, 20:24,  2 users,  load average: 0.04, 0.01, 0.00
>>>>>>              total       used       free     shared    buffers
>>>>>> cached
>>>>>> Mem:       2055108    1092176     962932          0      53156
>>>>>> 811864
>>>>>> -/+ buffers/cache:     227156    1827952
>>>>>> Swap:      4128760          0    4128760
>>>>>>  1.2 44308 perl
>>>>>>  0.0     0 aio/0
>>>>>>  0.0     0 async/mgr
>>>>>>  0.0     0 ata/0
>>>>>>  0.0     0 ata_aux
>>>>>>  0.0     0 bdi-default
>>>>>>  0.0     0 cpuset
>>>>>>  0.0     0 crypto/0
>>>>>>  0.0     0 events/0
>>>>>>  0.0     0 ext4-dio-unwrit
>>>>>>  0.0     0 flush-253:0
>>>>>>  0.0     0 jbd2/dm-0-8
>>>>>>  0.0     0 kacpi_hotplug
>>>>>>  0.0     0 kacpi_notify
>>>>>>  0.0     0 kacpid
>>>>>>  0.0     0 kauditd
>>>>>>  0.0     0 kblockd/0
>>>>>>  0.0     0 kdmflush
>>>>>>  0.0     0 khelper
>>>>>>  0.0     0 khubd
>>>>>>
>>>>>> Have to say i am at a total loss as there is no way the usage figures
>>>>>> are accurate.
>>>>>> If there is any other info i can supply that will be useful please
>>>>>> let me know.
>>>>>>
>>>>>> Regards
>>>>>>
>>>>>> Wayne
>>>>>>
>>>>>>
>>>>>> On 6 December 2011 12:03, Igor Homyakov <
>>>>>> [email protected]> wrote:
>>>>>>
>>>>>>> Hi Lawrence,
>>>>>>>
>>>>>>> Could you be a little bit more specific ?  Please provide information
>>>>>>> about you operation system, monit version on which the problem
>>>>>>> occurred and so on.
>>>>>>>
>>>>>>> Regards
>>>>>>> Igor Homyakov
>>>>>>>
>>>>>>> On Tue, Dec 6, 2011 at 15:35, Lawrence, Wayne
>>>>>>> <[email protected]> wrote:
>>>>>>> > Hi,
>>>>>>> >
>>>>>>> > I have a few CPU usage checks in my monitrc but it seems monit is
>>>>>>> > misreporting the usage.
>>>>>>> >
>>>>>>> > I have run several tests and it seems that monit is multiplying
>>>>>>> the actual
>>>>>>> > usage by 10.
>>>>>>> >
>>>>>>> > I ran a process with top running in another shell and CPU usage
>>>>>>> for the user
>>>>>>> > was never above 10% yet monit informed me that there was 100% cpu
>>>>>>> usage.
>>>>>>> >
>>>>>>> > I have tried various configurations including the one that came
>>>>>>> with the
>>>>>>> > default config for system cpu monitoring and all seem to
>>>>>>> demonstrate the
>>>>>>> > same issue.
>>>>>>> >
>>>>>>> > Any advice welcomed on this
>>>>>>> >
>>>>>>> > Regards
>>>>>>> >
>>>>>>> > Wayne Lawrence
>>>>>>>
>>>>>>
>>>>
>>>> --
>>>> To unsubscribe:
>>>> https://lists.nongnu.org/mailman/listinfo/monit-general
>>>>
>>>
>>>
>> --
>> To unsubscribe:
>> https://lists.nongnu.org/mailman/listinfo/monit-general
>>
>>
>>
>> --
>> To unsubscribe:
>> https://lists.nongnu.org/mailman/listinfo/monit-general
>>
>
> --
> To unsubscribe:
> https://lists.nongnu.org/mailman/listinfo/monit-general
>
>
>
> --
> To unsubscribe:
> https://lists.nongnu.org/mailman/listinfo/monit-general
>
--
To unsubscribe:
https://lists.nongnu.org/mailman/listinfo/monit-general

Reply via email to