Ad Widget

**kernbug** · 13-06-2018, 08:20

Originally posted by VR1387

I read:

https://www.zabbix.com/documentation/3.0/manual/appendix/items/vm.memory.size_params

I still do not understand the behavior

Thank you!

Hello

Please, show zabbix_get from Zabbix server for vm.memory.size[used] on Centos 7 and Centos 6, also for vm.memory.size[available] on CentOS 7 and CentOS 6.

And from console (6/7):
grep MemTotal /proc/meminfo | grep -E --only-matching '[[:digit:]]+'

**kloczek** · 13-06-2018, 14:38

Just FTR.
Is someone is using Linux with kernel 4.x memory management paradigm in Linux in those kernels slightly changed and now less important is free vs. used memory and sometimes more important is active vs. inactive.
This is why zabbix 4.0 will be able support new memory keys like:
- vm.memory.size[active]
- vm.memory.size[anon]
- vm.memory.size[inactive]
- vm.memory.size[slabs]

Patch for zabbix 3.4 with those new keys is possible to find on https://support.zabbix.com/browse/ZBX-13233

**VR1387** · 13-06-2018, 22:41

Originally posted by kernbug

Hello

Please, show zabbix_get from Zabbix server for vm.memory.size[used] on Centos 7 and Centos 6, also for vm.memory.size[available] on CentOS 7 and CentOS 6.

And from console (6/7):
grep MemTotal /proc/meminfo | grep -E --only-matching '[[:digit:]]+'

Hello

CentOS 7

This

htop

In Zabbix

In no case does it show me the memory usage of the server.
When I do the calculated item, it roughly gave me an approximate value but what I see is not the real use of the server.

Attached Files

**VR1387** · 13-06-2018, 22:42

Originally posted by kloczek

Just FTR.
Is someone is using Linux with kernel 4.x memory management paradigm in Linux in those kernels slightly changed and now less important is free vs. used memory and sometimes more important is active vs. inactive.
This is why zabbix 4.0 will be able support new memory keys like:
- vm.memory.size[active]
- vm.memory.size[anon]
- vm.memory.size[inactive]
- vm.memory.size[slabs]

Patch for zabbix 3.4 with those new keys is possible to find on https://support.zabbix.com/browse/ZBX-13233

centOS 6

This

htop

Zabbix

In no case does it show me the memory usage of the server.
When I do the calculated item, it roughly gave me an approximate value but what I see is not the real use of the server.

One of the closest values is the available memory % that good could do a trigger perfectly.

**kloczek** · 14-06-2018, 00:20

Originally posted by VR1387

One of the closest values is the available memory % that good could do a trigger perfectly.

Code:

$ grep -i active /proc/meminfo ; uname -a
Active:          5565924 kB
Inactive:        1254940 kB
Active(anon):    4869884 kB
Inactive(anon):   618044 kB
Active(file):     696040 kB
Inactive(file):   636896 kB
Linux domek 4.18.0-0.rc0.git2.1.fc29.x86_64 #1 SMP Tue Jun 5 20:01:21 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux

Embedding triggest about memory monitoring in kind of base OS type template is really wrong.
Why?

Here is a bit longer answer ..
Because in some exact scenarios application running on exact system may be allocating ALL possible to allocate memory leaving few small scraps and looking on % free/avail memory triggers you may have impression that something is wrong on such system because application is suffers by "memory starvation". In such case you want to have completely opposite type of alarm that on exact system. Such alarm could be used to check application memory allocation settings or to decrease system memory to reduce cost of running platform for exact application.

Typical case it is DB backend.
Oracle DB, PostgreSQL, MySQL and other are allocating exact amount of memory and those applications usually opens all files in O_DIRECT. Those applications/services are organizing in allocated memory own caching of the database content. In exact scenarios this may be far more effective than using the same memory as buffer cache.
Even only on the area like DB backends cases situation may be opposite.
For example when I'm using SQL engines running on top of the Solaris with ZFS beneath I'm giving for example for MySQL innodb pool only memory to have enough hits/misses ratio into indexes. ZFS caching layer ARC + optional L2ARC and especially with enabled ZFS transparren compression provides way better caching in ARC than MySQL in innodb pool. Why? mostly because ARC keeps in cache compressed content available for the application over VFS layer.
In such case physical memory used by ARC works so effective like its physical size multiplied by avg compression ration. In other words if your system is using 100-200GB of ARC and data stored in ARC is possible to compress with 10x compression ratio? Such 100GB works like 1TB RAM but without compression.
If someone is something have been scratching own head asking them self "why today when MySQL or PostgreSQL is for free some people are ready to pay huge licensing costs for Oracle DB?" answer is: because this engine with columnar compression is able to provide in case of some exact DB contents 10-30x compression ratio .. level unreachable for any MySQL or PostgreSQL. Try to only think what is the HW price difference between typical pizza box HW sometimes even with 1TB memory and other HW which would be able provide for single OS image/instance 20-30TB of RAM?

Other typical representation of those applications which allocates usually as much as it is only possible is JVM.
Usually as long as java application has nothing to do with local FS interaction someone will be allocation almost 100% possible to allocate memory leaving only memory allowing to login over single ssh session.
In case of OS layer memory monitoring definitely you would want to be alarmed that OS memory is opverspect or JVM settings are wrong. Isn't it?

Completely opposite case are as well possible.
You may have application or service running on exact system which completely relays on OS layer buffer cache and it allocates only small chunk of available memory.
Typical case it is NFS server.
Someone who don't know what kind of application is running on such system may have impression that exact system has hugely overspect memory, ad as result of the "to much free memory" alarm may be even trying raise change request to reduce memory on exact system or requesting migrate exact application/service to the physical HW with less memory.
Of course NFS server on Solaris using ZFS resources is integrated with ZFS caching so in this case it will be only using ZFS ARC/L2ARC so looking on buffer cache metrics provided over kstat you may be looking on wrong metrics.
On Linux you may be interested SLAB allocator stats but on Solaris you shopuld be looking on ARC/L2ARC monitoring data.

htop, top or other like vmstat, iostat are nothing more than just tools providing some data which after sampling needs to analysed with other details which those tools cannot even reach or add to generated view of those data.

And yet another small exampler from my laptop:

Code:

$ wc -l /proc/meminfo
48 /proc/meminfo

As you see latest Linux provides really reach set of metrics about memory usage.
Only one of the line from above:

Code:

$ grep -i slab /proc/meminfo
Slab:             708820 kB

Could be unrolled to:

Code:

# wc -l /proc/slabinfo
150 /proc/slabinfo

Did you try to ever use slabtop command on Linux?

Do you see now how far from accurate could be sometimes using only free/avail memory metrics?

Ad Widget

imetm key zabbix 3.0 vm.memory.size [used

imetm key zabbix 3.0 vm.memory.size [used

Comment

Comment

Comment

Comment

Comment