Is necesary cache memoty?

[Jimmy_13]

I have a SuperDome,32 GB RAM, whith a storage XP128.
In the begining the cache eat 80% of memory, whe modified the kernel whit dbc_max_pct=2 the use of memory reduce to 18 %.
I need cache memory? or theXP128 have the cache for i/o?
If the cahe memory ins't necesary how i can turn off o reduce to min?

Thanks

[Krishna Prasad]

I guess you could change the max_pct = 0, or set the number of buffer pages to a fixed very small amount. I wouldn't set it all to zero unless even your root/boot disk is in the XP128. The OS will use the buffer cache. However, since your XP will be doing the majority of the I/O you have the right idea to limit the buffer cache as much as possible.

[A. Clay Stephenson]

I wouldn't turn it off completely; no matter how fast the disk array cache is there is still i/o overhead involved in getting to it so your box will benefit from UNIX buffer cache especially in the case where many processes are using the same files (e.g shared libraries). If this is an 11.0 box, I would limit the cache to about 800MB and if it's 11.11 somewhere around 1200MB would be a good maximum. You can hard-set the buffer cache to a fixed size by setting bufpages to a non-zero value. Bufpages is set in 4K pages so 204800 would be 800MB.

[Tim Adamson_1]

I wouldn't disable it all together. Leave it at the current value (2) and monitor the situation. You will probably have disks that are NOT on the XP and hence will still be using buffer cache.


Tim.

[James R. Ferguson]

Hi Jimmy:

If you are running Oracle, you can also reduce the buffer cache's use (i.e. bypass it) by mounting your Oracle data (not rollback or redo log) filesystems with:

delaylog,mincache=direct,convosync=direct,nodatainlog

You need Online JFS to do this, but you may find that your performance will improve.

Regards!

...JRF...

[Todd McDaniel_1]

I definitely agree with JRF...

I have a Superdome 48way(750MHZ) with 72GB memory, with several EMC Symms attached... I left the local buffer cache intact and have all my DB filesystems with the options:

mincache=direct,convosync=direct... along with largefiles,delaylog,nodatainlog


My production DB is 1.2TB and my Superdome could not function if I relied solely on the OS buffer cache...