large buffer cache

[ASO FINANCE]

Hi all,
I have a customer with a Superdome with 1 hard partition of 32 CPU and 64 Gb RAM.
we have a LARGE Buffer cache of 6,4 Gb and I think that it's a negative factor for the system.Customer have Oracle 9.2 in this partition and now he see some performance problem of I/O.buffer cache read Hit it's good (about 90%) but wrute cache Hit it's very poor (40-50%).
It's a good idea to decrease the buffer cache about 1Gb max?
I need to use convosync=direct,mincache=direct to improve Oracle performance?
have you some document about the right size of Buffer cache?
thank you very much.
Max

[Jeff Schussele]

Hi Max,

Yes - a LARGE OS buffer cache can impede performance.
Yes - Oracle performs better with a large SGA & letting that buffer the data & bypassing the OS buffer cache with convosync=direct & mincache=direct
BUT only do this with the data & indices. The redo & archive FSs should use the buffer cache.
We find between 400 - 800 MB to be the sweet spot for buffer cache size.

HTH,
Jeff

[RAC_1]

The read cache is low, so there is certainly a room to decrease it and watch read and write hit ratio.

for putting convosync=direct,mincache=direct, you need to have online jfs. and it is just a metter of unmounting systems and mounting them with those options. You then compare the results.

Anil

[Fabio Ettore]

Ciao,

I suggest you to check these very good threads about this argument (oracle and dbc_max_pct):

http://forums1.itrc.hp.com/service/forums/questionanswer.do?admit=716493758+1085494754762+28353475&threadId=467288

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=588393

HTH.

Best regards,
Ettore

[Jeff Schussele]

Hi RAC,

unmounting is not necessary if you have OnLine JFS 3.3 or higher.
You can use the -o remount option to change it on the fly.

Rgds,
Jeff

[RAC_1]

Thanks Jeff. I had completely forgotten -o remount option.

That's why I love this forum.

Anil

[Jeff Schussele]

Anytime - my friend.
I love it as well.

Rgds,
Jeff

[ASO FINANCE]

thank you,
I made some test with different option:
my buffer cache now is 800Mb and Oracle FS are mounted with option mincache=direct,convosync=direct and with Oracle people we saw the new performance of the DB and the O.S. System too.
However,my write Hit now is higher 80% and seems that the DB disks write a lot.
But Oracle people saw that it seems that there is more queue in read.
I think that it must be for the bypassing of buffer cache that in write it's very good but in read have some problem of speed.
there is a parameters that permit to a FS to use buffer cache only in read?
I think that it don't exist but my customer ask me..every minute!!