You have mentioned this towards the end of the posting: "You can also improve redo performance by making sure that the redo log files reside on fast disks with no I/O taking place on them other than the LGWR writes."
An average OLTP system can have about three online logs. Let us also assume each log is mirrored in 2 disks at least. I am assuming that is a fairly common scenario where RAIDs are used and where SANs are not involved.
If we want no I/O other than the log writes on the disks we may have to have separate disks for each log file (we need about 3x2=6 in the above scenario). Is that a solution adopted commonly? This may mean buying low capacity (few GBs), fast hard disks for log files alone? How do shops make decisions on that? Also, how does it work in the case of SANs?
Your analysis is correct. The ideal setup for redo log files is disks that are dedicated to the logs. With the increasing use of SANs and other array technologies, though, this is increasingly harder to achieve. In many cases, placing redo logs on virtualized storage does provide acceptable performance, but you should monitor redo performance by observing, for example, waits for redo writes. If these start to become significant, you can then use the information as justification for adding a mirrored pair of directly-attached disks to your database server. If you can't get relatively small disks for this purpose, you can use the excess capacity of the drives to store database backups; since the backup activity is presumably limited to a small time window, it won't affect redo performance at all when your database is most active.