I decided to perform an experiment. Before I share the results, I'll explain how I set up the experiment. Take a deep breath, and continue.

The server is a single Intel CPU/core Dell box with 512MB of memory, running Oracle Database 11g Release 1, with a 256MB of shared pool memory, a buffer cache of 4MB (yes, I know it's small, but it forces a lot of IO), and running on Oracle's unbreakable Linux 5. Each of the two main tests (IMU enabled and IMU disabled) was run nine times. The worst time was removed from each series, resulting in eight samples. Before each of the two tests, the instance was recycled, and the DML load was started and left to stabilize for about 5 minutes. Then a heavy read-consistent query was run four times. CPU and wall clock time statistics were collected before and after each query; that is, the wall clock time is the difference between the time at the end of the last query subtracted from the time before the first query.

For this particular experiment, there was a statistically significant 21% CPU time reduction when using IMU. However, there was a statistically insignificant 6.5% wall clock time reduction when using IMU. This means, for this particular experiment, there is not quite enough wall clock time difference to indicate using IMU made an improvement.