Buffer-cache

[Jdamian]

A V-class system with HP-UX 11.00 swapped a little. I cannot understand why.

The physical memory is 3 GB, 2 of them is available for user processes.

Glance reported that the memory space is used by user processes and by buffer-cache.

My little knowledge about buffer-cache made me to suppose that if a process need more memory the buffer-cache will be reduced in order to free mem pages.

[Duane Gorder · Accepted solution]

My experience has been that buffer cache does not release memory nearly fast enough to satisfy my applications. I have decreased my dbc_max_pct to ten percent. This would be 200MB on my smallest system. This is enough for the applications that I am running. Any other application would have to be monitored to make sure performance does not degrade.

[Steven Sim Kok Leong]

Hi,

There was one occasion when incompatibility between OAS and the Oracle database caused each Oracle user process to take up 150 MB of memory. In a matter of minutes, the entire memory was utilised and there was deactivation of processes. After the OAS was fixed, each user process took up 20 MB of memory and everything resumed back to normal.

Deactivation only occurs when your process is too big to fit into memory. If you have sufficient memory, you should also see minimal paging out. Instead, you should see a lot of page ins most of the time.

In your case, perhaps you want to use glance or top to identify your memory-intensive processes and your available free memory. It could be a memory leak or a erratic memory consumption (like in my case previously).

Hope this helps. Regards.

Steven Sim Kok Leong
Brainbench MVP for Unix Admin
http://www.brainbench.com

[Byron Myers]

Was the swapping on device or "memory". "swapinfo -tm" will show this. If you did indeed swap to device, then investigation is warranted. If you swapped to "memory" this is normal.

[Philip Chan_1]

I thought that when the OS tries to release buffer cache the first thing to be done is to flush the cache content to disk, hence this will trigger certain physical I/O that slow down the system. So, I also believe that the
dbc_max_pct param should not be lefted at its default (ie. 50%). My gut feel is that 20% percent would be more appropriate.

~Philip