BLOB Performance

I'm running the following benchmark transaction on a Sun 4500 with 8 400M processers. I have 4 mirrored drives where I have (in general) data, indexes, logs, and 1 heavily hit table with BLOBs separated. My page size is 8192 and I have 128M in db_block_buffers. (I could give it more, but I'm trying to simulate the data-cache ratio of a VLDB.) Log size is 10M.

4 tables are inserted directly
1 table is inserted by a trigger
2 tables includes BLOB's of 40K average size.

Here is my question.

Sending the above transactions on 1 client thread and Oracle connection yields a Tps=16. Two threads and connections, both pumping the above Tx as fast as it can yields a Tps=22. When I bump the comnnections to 3...the performance collapses to Tps=11. (?) I'm not a 'utlbstat' expert, but nothing stood out in terms of waits, hit/ratios, etc.

Any clues as to what causes my performance to drop so rapidly?

--
Ron Boggs
Trade Harbor, Inc.
St. Louis, MO


Sent via Deja.com http://www.deja.com/
Before you buy.