[Oracle-l] Linux Memory Accounting

Does anyone have a fairly accurate method for determining
how much memory is actually being used on a Linux box?

The linux in question is RedHat Enterprise Server 4.

...yes, free(1):

[root_at_tmr6s14 ~]# free

total used free shared buffers
cached
Mem: 32752464 1904580 30847884 0 65444
1523444

-/+ buffers/cache: 315692 32436772
Swap: 2040244 0 2040244

...you rephrased the question after posting to the oaktable,
phrased this way took a lot less typing to answer :)

Thanks, though I think you knew what I meant. :)

To be more precise in what I am asking:

According to the ps man page:

rss RSS resident set size, the non-swapped physical
memory that a task has used (in kiloBytes).

(alias rssize, rsz).

What's interesting is that summing RSS for all Oracle processes
on the server in question (RH ES 4 64 Bit) yield 25 gig of RAM.

The server has only 12 gig of RAM.

Here's the ps command:

ps -fywluoracle | awk '{ mem=mem+$8 } END{ mem=mem*1024; print "bytes: ",
mem}'

My question is: Does someone here know what is really being reported by
RSS?

Or is it just untrustworthy?

Shared memory shows up in the RSS of each process....so I suppose you
could do complicated math surrounding subtracting out the SGA for all
processes that attach therein, and then add it up....


Matt

From: oracle-l-bounce_at_freelists.org
On Behalf Of Jared Still

Sent: Monday, October 02, 2006 5:04 PM
To: kevinc_at_polyserve.com
Cc: Oracle-L Freelists
Subject: Re: Linux Memory Accounting

Thanks, though I think you knew what I meant. :)

To be more precise in what I am asking:
--------------------------------------

According to the ps man page:

rss RSS resident set size, the non-swapped
physical memory that a task has used (in kiloBytes).
(alias rssize, rsz).

What's interesting is that summing RSS for all Oracle processes
on the server in question (RH ES 4 64 Bit) yield 25 gig of RAM.

The server has only 12 gig of RAM.

Here's the ps command:

ps -fywluoracle | awk '{ mem=mem+$8 } END{ mem=mem*1024; print

"bytes: ", mem}'


My question is: Does someone here know what is really being
reported by RSS?


Or is it just untrustworthy?

$cat /proc/meminfo

Thanks, but that cannot be used to determine how much memory Oracle is
using.

Please see my previous reply on this.

Jared

Do you have a script/process that does that?

I'm not into reinventing the wheel here if I can avoid it.

Would be sweet if linux actually knew how to properly account for its
memory.

Jared

Try MetaLink note 269426.1

Hemant
Hemant K Chitale
http://web.singnet.com.sg/~hkchital

That note refers to statm|stat|status pmap, etc.

Those are all inaccurate for real memory accounting.

jared,
processing RSS is a wrong way of calculating memory usage.
It's alrights for private memory/PGA, but oracle uses shared memory, in
which the SGA resides.
That memory is not bound to a specific process, rather a linux/unix system
keeps track of the processes which access them, so current PID and creater
pid are kept.
the RSS is calculated of private memory + touched shared memory pages
this means accumulating all RSS amount for all database processes, the
chance you calculate memory twice is fairly big, and gets higher once the
instance is longer alive (the background processes are able to touch more
shared memory)

this means it's impossible to calculate database memory usage in a simple
way by examining the processes at OS level.

it's also worth to note that shared memory usage display is turned off by
default in linux (at least in the recent 2.4 kernels and the RHEL3 and
4 2.6kernels) in /proc/meminfo. A way to get linux display shared
memory pages is
(of course) ipcs -ma, and also with (as root):
echo m > /proc/sysrq-trigger; dmesg | grep shared

so I'd go with kevin's recommandation to use some tool to look at global OS
level to see memory usage. besides the omitted shared memory, it's accurate.

frits

Hi Frits,

Comments inline:
Yes, I've discovered that.

so I'd go with kevin's recommandation to use some tool to look at global OS
I've cobbled together a script to scan the /proc//maps files for
a process owner, and add up the memory, hopefully with a minimum
of double counting.

The results for running it on 2 servers are approximately what I
would expect to see. This includes shared memory.

http://jaredstill.com/downloads/proc_mem_used.tgz

Comments, suggestions and/or improvments are welcomed.

Hey Jared,


I'll have to check this out (may help with new ERP go-live), but you
seem to imply that this accounting issue is only an issue on Linux.
This isn't the case for HP-UX, Solaris, and AIX?


I can't say I've treated it any other way, although I can only recall
needing to look at this info once before...


Rich

From: oracle-l-bounce_at_freelists.org
On Behalf Of Jared Still
Sent: Tuesday, October 03, 2006 4:58 PM
To: Frits Hoogland
Cc: Oracle-L Freelists
Subject: Re: Linux Memory Accounting

Hi Frits,

Comments inline:

On 10/3/06, Frits Hoogland wrote:

this means it's impossible to calculate database memory usage in
a simple way by examining the processes at OS level.

Yes, I've discovered that.

so I'd go with kevin's recommandation to use some tool to look
at global OS level to see memory usage. besides the omitted shared
memory, it's accurate.



I've cobbled together a script to scan the /proc//maps files for
a process owner, and add up the memory, hopefully with a minimum
of double counting.

The results for running it on 2 servers are approximately what I
would expect to see. This includes shared memory.

http://jaredstill.com/downloads/proc_mem_used.tgz

Comments, suggestions and/or improvments are welcomed.

Jared,

I am working on a presentation that will answer these exact questions.
I have the abstract here:

http://www.pythian.com/blogs/255/the-answer-to-free-memory-swap-oracle-and-everything