The argument goes that raw devices are faster because they bypass the overhead of the OS buffer cache that is associated with filesystems. Also, on some platforms kernalized asynchrous IO can only be used with raw devices and not filesystems. The argument also goes that raw devices are harder to manage than filesystems.

However, there is a lot of debate about how true this is with todays technology. For example:

1) Some platforms allow filesystems to be mounted in a direct mode that also bypass OS buffering
2) Modern logical volume managers mean that raw devices aren't really that hard to manage
3) Newer extent based filesystems (e.g. vxfs) mean the performance difference between raw and filesystems is negligble.

To a large extent it probably depends on your platform and needs to be benchmarked. Of course, if you're using Parallel Server or RAC you must use raw devices.

If you're considering migrating to raw devices you might want to read this word of warning. It doesn't always mean free performance improvements.

HTH

Austin