While the percentage of recursive SQL was the same with mutexes enabled and disabled, when latches were used, the total CPU consumption was nearly twice as much! In fact, look closely and compare the time waited for library cache-related events in Figures 7-11 and 7-12. Once again, mutexes are shown to be more efficient than latches.

Figure 7-12. Shown is an example of severe library cache contention when mutexes are enabled. The same exact load and reporting interval was used as in Figure 7-11.

As was shown in Figure 7-9, Oracle is very particular about what it considers a similar SQL statement. Every statement must be parsed, and if the cursor is not found in the library cache, the cursor must be completely built (a hard parse). Hard parsing requires library cache-related latches and locks, so if hard parsing becomes so intense the related wait events are driven to the top of our reports, we will look for ways to create similar SQL statements. Oracle provides two powerful methods to do this.