| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
 |
 | |||
| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
|
| |||
Home -> Community -> Usenet -> c.d.o.server -> Possible to turn off file system cache?
Hi There,
I'm working on a Solaris 2.5.1 machine which appears to be memory constrained.
When I run sar -r 5 5 I see this:
texas% sar -r 5 5
SunOS texas 5.5.1 Generic_103640-04 sun4u 04/19/97
22:42:39 freemem freeswap 22:42:44 2085 2896453 22:42:49 2110 2896128 22:42:54 2162 2908621 22:42:59 2154 2896875 22:43:04 2160 2902493 Average 2134 2900106
It only has 2134 pages of freemem.
If I run whatami I see that it has 1gb of ram:
DATE: Sat Apr 19 23:21:14 PDT 1997 <MOVE REQUEST INFO>
USER: oracle <LOGIN = oracle>
MODEL: SUNW,Ultra-Enterprise <MACHINE MODEL = SUNW,Ultra-Enterprise>
CPU: SUNW,UltraSPARC
CPU: SUNW,UltraSPARC
CPU: SUNW,UltraSPARC
CPU: SUNW,UltraSPARC
SunOS RELEASE: 5.5.1
MEMORY: 1024MB
SWAP: 1686.8MB total, 269.2MB used, 1417.5MB available
LOAD AVERAGE: 0.12, 0.14, 0.14
If I look at shared memory used by Oracle I see this:
SVRMGR> show sga
Total System Global Area 216062460 bytes
So, it looks like about 20% of the memory is going to Oracle as shared memory (aka SGA in oracle parlance).
If I look at processes I see a few large oracle processes which look like this:
ps -e -o user,pmem,vsz,args | grep oracle
USER %MEM VSZ COMMAND^
oracle 0.2 3032 /u01/app/oracle/product/7.2.2/bin/tnslsnr LISTENER -inherit oracle 20.7 223584 ora_dbwr_proddb oracle 20.7 222952 ora_arch_proddb oracle 21.0 222792 ora_smon_proddb oracle 20.8 222752 ora_reco_proddb oracle 21.0 222736 ora_pmon_proddb oracle 20.7 222912 ora_lgwr_proddb oracle 0.1 1088 csh oracle 0.2 1568 -ksh oracle 0.1 1088 csh oracle 0.1 832 grep oracle oracle 0.2 1576 -ksh
Inspection of the process listing above reveals a question,
"When ps looks at a particular oracle process, is the shared memory which the process is attached to count towards the memory size of that process?"
For example, if we look at the listing for ora_dbwr_proddb, we see that ps believes the size of the process to be 223584 kb.
This is a huge memory size for a process. I can only make sense of it if I conclude that ps is adding shared memory to private memory in the display but that the process is not really 223584 kb. If I want to find the true size of the process I need to subtract 216062460 bytes from 223584 kb to come up with 12586 kb. This is still a large size for a process but we are talking about an Oracle process (Oracle software has a rep for being a resource hog).
Actually, now that I think about it, one other scary scenario comes
to mind:
If a process connects to shared memory, that process makes a copy
of that shared memory. Thus, when ps sees such a process it sees
a truely huge process.
This second scenario is so scary, in fact, that I'll ignore it and concentrate on scenario # 1.
So, I decide to run with this wild guess. If I want to find the true size of all the processes running, I need to subtract out the size of the shared memory of all the giant oracle processes.
The following script is helpful:
#! /bin/sh
echo 0>/tmp/dcme
ps -e -o vsz | /bin/awk '{print $0"+" }'|grep -v VSZ>>/tmp/dcme
echo p>>/tmp/dcme
echo q>>/tmp/dcme
dc /tmp/dcme
################################################
When I run it I get the following number:
1439960
So, ps reports that the total memory consumption by processes on this machine is 1,439,960 kb (which is outrageous).
Next, I inspect /tmp/dcme and notice 6 entries for the huge oracle processes. So, to adjust the number I multiply 216062460 bytes by 6 and subtract the result from 1,439,960 kb to get:
1,439,960 kb - (216062460*6/1024) kb = 173970 kb
So, it appears that at this moment, all of the process on this machine consume only 173,970 kb which is about 17 % of the 1gb of ram.
Now the arithmetic gets simple...
Shared memory consumes 20% of the ram
Processes consume 17% of the ram
WHERE is the other 63% of the ram?
I suspect the main clue to the answer lies in the one of my dusty UNIX books. I think I remember reading that UNIX, like oracle likes to set up buffers in memory to cache files which are active.
Could it be that Solaris is devoting 63% of the ram to cache files which have been opened by oracle?
On this machine I happen to know that all of the oracle data files are large ufs files rather than raw devices.
Does Solaris have a shell command or gui tool which will tell me how much ram is devoted to caching the file system?
Does Solaris have a shell command or gui tool which I can use to tell Solaris to not cache certain files?
-Dan
 |
 |