cbc latch assignment on db_cache resize

Now Oracle has a dynamically resizeable buffer cache, this obviously presents some interesting properties for how the buffer cache and associated objects get rearranged when the size is changed.

In particular, does anybody know what Oracle does with cache buffers chains latches when the buffer cache is resized?

As I see it, there are (at least) two options:

1a (on a shrink) Remove a proportion of the latches and rehash buffers across the remaining latches.

1b (on a grow) Add some more latches, and rehash all buffers. (Sounds expensive...)

2 Have enough hash latches for all possible buffers in sga_max_size, and create or remove buffers keeping a constant number of latches as db_cache_size changes.

I ask because I was running a test, after the test I dynamically reduced db_cache_size to 80% of its former value (100GB to 80GB), then re-ran the test. It ran fine for a while, then wrapped itself around cache buffers chains latch wait (89% latch wait, all cbc) for two distinct periods of several minutes each. After shutting down the database and restarting with a SGA of the smaller size, the test ran to completion twice without problem.

I'm wondering if the massive cbc latch contention was caused by some artifact of resizing the buffer cache, or if it was pure bad luck and could happen at any point. The two latch children that were suffering the waits protected many blocks spread all across a couple of fairly large indexes (320m rows each) that were being hit quite hard.

Any ideas?

(Oracle 9.2.0.1, AIX 5L)

-- 
Andrew Mobbs - http://www.chiark.greenend.org.uk/~andrewm/