Cach Size

[George Nikoloudis_1]

Hello to all
I would like to ask:
I have a V Class System 8 GB Memory 8 Processors. I run Oracle 8.1.7 (8i) with SGA 1GB. I have FC10 for storage. As I know these hard disks do not have any cache.

Would it be good ides to use 1GB of my memory for cahcing between SGA and FC10?

How can I check the size of the cache I have now, and how can I change the cache size.

Thanks a lot
George Nikoloudis

[A. Clay Stephenson · Accepted solution]

Hi George:

To find those values you can do a kmtune and look for 3 values: dbc_max_pct,dbc_min_pct, and bufpages. If bufpages is non-zero, multiply that value by 4K and that is your UNIX buffer cache. If bufpages = 0, the buffer cache varies between dbc_min_pct and dbc_max_pct as percentages of your total memory. In most case, 1GB is way too much buffer cache and will actually degrade performance. I would set bufpages to about 80000 (320 MB and that is generous) and measure. Bear in mind, if you are using raw/io buffer cache does not matter or if your are using the OnlineJFS mount options convosync=direct,mincache=direct which also bypass the buffers. I do suggest that you do not use dynamic buffer cache and set bufpages to some non-zero value.

[George Nikoloudis_1]

I have confgure the ON LINE JFS for direct acces. However I did not see any difference. What I was thinging is to get this buffer cache on the KERNEL to 320 MB and bring back the OnLine JFS parameters to default.

What do you thing?

[Roger Baptiste]

Hi,

You can check your current cache size/usage by going into glance and use the "t" option .
You can check your Cache hit rate by doing
sar -b
-> it should show readcache and writecache hit rate. Values of 100% and 90% respectively means your cache size is good enough!

I agree with Clay, 1Gb is toooo much for a buffer setting. The thumrule is a value around 10% of the system memory upto a maximum of 400-600mb (in case you have large memory).

<>

Are you having a performance issue with the database?? If so, you would need to make sure where the problem is. Do you have pvlinks set up to the disks. It would help in load balancing. Are your lv stripe sizes set correctly? What is the PE extent size?
What is the FS block size?
The dba should also make sure his sga is configured correctly.

<>

320MB may be a bit less, since your RAID does not have internal cache (as per your post), unlike the EMC or XP arrays. You could use dynamic max of 7% and min of 5%.
I don't use the JFS parameters on any of my Oracle systems.

HTH
raj

[A. Clay Stephenson]

George:

Here is what I have found to work well and to indeed be as good as raw/io.

Datafiles/Indices convosync=direct,mincache=direct,nodatainlog,delaylog,rw

Archive/Redo Logs nodatainlog,delaylog,rw

If you don't set nodatainlog, you are limited to 1MB chunks before AND the data has to written twice. So set this option before going back to cooked i/o. On 11.11 I have found that often cooked i/o is actually better than raw for Oracle in some cases - but again the real answer is to measure for yourself. This does assume very generous buffers in the SGA where Oracle likes to do its buffering.

By the way, it is considered polite to assign point to those who try to assist you; your points assignment has been terrible.

Clay