[users at i-scream] libstatgrab on AIX: data mismatch between saidar and topas

Anderson Carlos Trindade anderson.trindade at optimode.com.br
Tue May 24 13:51:22 BST 2016


Hi Jens,

So, the point is:

	- libstatgrab is reporting the physical CPU usage. If libstatgrab shows something around 65% of Idle, it means that 65% of all physical resources are Idle.

	- on the other hand, the sample code is reporting the LPAR usage. If the sample code shows something around 20% of Idle, it means that LPAR has just 20% of the CPU dedicated to LPAR is available to LPAR usage

Is this understanding correct?

Considering I have one application running inside a LPAR and this application is consuming almost all CPU dedicated to LPAR (around 80%) but the physical host is using just 35% of CPU, If libstatgrab returns the physical usage, I can’t see from the libstatgrab perspective that the LPAR is almost 100% of CPU usage. Is that correct?




> Em 24 de mai de 2016, à(s) 06:11, Jens Rehsack <rehsack at gmail.com> escreveu:
> 
> Hi Anderson,
> 
> the example is very explicit about the measurement - it normalizes the values when lparstats.type.b.shared_enabled - libstatgrab doesn't.
> libstatgrab reports the physical cpu measure - which can lead to misinterpretion for shared resources (which is up to our knowledge always the case when physical resources are shared on a best effort way). So we decided against that (similar for zones (Solaris), Jails (BSD) and Containers (Linux)) until we find a tuit to analyze all available technologies and a reasonable way to deal with them.
> 
> Thanks for remind me :)
> 
> Best regards,
> Jens
> 
>> Am 23.05.2016 um 18:54 schrieb Anderson Carlos Trindade <anderson.trindade at optimode.com.br>:
>> 
>> Hi Jens,
>> 
>> Thank you for reply!
>> 
>> As far as I know, topas seems to be an AIX utility (https://www.ibm.com/support/knowledgecenter/#!/ssw_aix_71/com.ibm.aix.cmds5/topas.htm), but I can’t tel you where the data displayed by topas is coming from.
>> 
>> But let’s forget topas for a moment.
>> 
>> I got a sample code from IBM site (https://www.ibm.com/support/knowledgecenter/#!/ssw_aix_53/com.ibm.aix.prftools/doc/prftools/prftools07.htm%23wq407), which uses perfstat to report cpu usage statistics.
>> 
>> Then, I compiled this sample code and run in parallel to saidar, each one in a separated SSH session. While saidar is reporting around 80% of Idle time and 10% of user time, the sample code above (based on perfstat) is reporting something around 35% of idle time and 60% of user mode usage. I recorded a screenshot and I can share with you If you prefer.
>> 
>> Considering that saidar and the sample code above are getting data from the same source, why are these statistics so different?
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>>> Em 23 de mai de 2016, à(s) 13:02, Jens Rehsack <rehsack at gmail.com> escreveu:
>>> 
>>> Hi,
>>> 
>>>> Am 23.05.2016 um 17:02 schrieb Anderson Carlos Trindade <anderson.trindade at optimode.com.br>:
>>>> 
>>>> Hello List,
>>>> 
>>>> I’m trying to understanding some differences between data reported by AIX utility topas and saidar.
>>>> 
>>>> On a given moment (almost in the same second), saidar report the following CPU usage:
>>>> 
>>>> CPU Idle: 88,28%
>>>> CPU system: 4,72%
>>>> CPU User: 7,00%
>>>> 
>>>> but topas report the following usage:
>>>> 
>>>> %Idle 35,8%
>>>> %Kern 3,5%
>>>> %User: 60,5%
>>>> %Wait 0,2%
>>>> 
>>>> It seems that both utilities are using different sources of data, since the usage reported is very different.
>>>> Please, could you help me to understand where these differences are coming from?
>>> 
>>> Well, I don't know where topas is fetching it's data from - and where your topas comes from (AIX Linux Tools? 3rd party repo?) ....
>>> 
>>> As you can see here https://github.com/i-scream/libstatgrab/blob/master/src/libstatgrab/cpu_stats.c#L162, libstatgrab is using perfstat - the IBM recommendation and the same source used by nmon.
>>> See https://www.ibm.com/support/knowledgecenter/ssw_aix_53/com.ibm.aix.prftools/doc/prftools/prftools07.htm%23wq407 for more details about libperfstat.
>>> 
>>> The reason for enhancing libstatgrab by a former customer was the poor data quality of GNU tools on Unices (HP-UX, AIX, Solaris).
>>> When I'm in doubt, I trust libstatgrab more than all GNU tools together >:-)
>>> 
>>>> my apologies in advance, because I'm very new on AIX world
>>>> 
>>>> This is a LPAR with 4 CPU’s
>>> 
>>> Best regards
>>> --
>>> Jens Rehsack - rehsack at gmail.com
>> 
> 
> --
> Jens Rehsack - rehsack at gmail.com
> 




More information about the users mailing list