Hi !
I want to eliminate the pagecache as source of our problem, that's the
reason why i want to limit it to a reasonable value for a DB-Machine (for
example 1GB).
We have a 64bit machine with 32GB Memory. The SGA is 10G with Oracle
10.2.0.1 and we are watching unexpected results in long running batches.
- when the instance is cold started (but the machine is warm), the test runs
fine with very good performance
- when the instance is warm and we start the test, after 2 hours performance
is degrading und after another 1 hour the performance seems to be ok again.
- the test involves 2 large tables (about 8-12GB), consuming the whole
SGA-Buffercache, because it is a sequential scan on all rows via Index (not
a fts).
I worry about the page-cache, because we see a pagecache (/proc/meminfo ->
Cached) of 31GB (!?!?!) while the instance is running (and the test) and the
SGA-BufferCache is nearly full. So could it be the interference between the
pagecache and the oracle-SGA Buffer-Cache? As far as the buffercache is
filled, all row-accesses have to make a lookup through a huge table of rows
and their latches. And: Memory-Performance is finite! Even on new fast 64bit
machines. And it makes no sense to make a double-buffering...
I fear, that parts of the SGA are swapped out to increase the page-cache and
so if i limit the page-cache to approx. 1GB there is one parameter
eliminated in our problem.
We are currently not using HugeTBL.
About other suggestions:
- Async-IO is on
- filesystemio_options are set the asynch
- Swappiness is on default (60)
- We see swapping from time to time
- io-scheduler is cfg (is there any way to determine this through proc
instead of dmesg?)
- iosched/quantum is set to 4 on all devices
We will try to modify the swappiness and the io-scheduler...let's see any
changes..
Any experiences with "lock_sga=true"? Will it help to stick the sga in
memory and to avoid swapping out parts of the SGA on Linux? Would it help to
have larger pages (page-clustering or hugetbl or similar) to avoid large
memory-operations in kernel while checking the pagecache etc.?
Thanks for all the quick help...stay tuned ;-)
Bye
Frank
_____
Von: Fabrizio Magni [mailto:fabrizio.magni@(protected)]
Gesendet: Dienstag, 8. August 2006 08:53
An: Frank Westheider
Cc: suse-oracle@(protected)
Betreff: Re: [suse-oracle] Hot to limit the pagecache in SLES9
As far as I know pagecache was a redhat proc entry. But I don't have a SLES8
around anymore where to check.
In theory the parameter has (and had) few sense. Cached pages are reused
when needed. Why to waste unused memory?
If you fear to swap there is a swappiness parameter.
Regards
Fabrizio
On 8/7/06, Frank Westheider < <mailto:frank.westheider@(protected)>
frank.westheider@(protected):
Hi !
I want to limit the pagecache in SLES9-2.6 Kernel.
In 2.4 there was a parameter vm.pagecache or /proc/sys/vm/pagecache..
This parameter is gone in 2.6..
Is there any way to limit the pagecache in 2.6?
We have a serious Oracle-Cache-Performance issue and i want to wipe out the
system-cache to be sure, that it is an oracle problem, not a
cache-interaction-problem.
Best regards
Frank
======================
Frank Westheider
Westheider IT-Service
--
To unsubscribe, email: suse-oracle-unsubscribe@(protected)
For additional commands, email: suse-oracle-help@(protected)
Please see http://www.suse.com/oracle/ before posting
