A very simple and straightforward illustration of this relates to latch contention. As discussed in Chapter 3, the latch wait event is process latch sleep time, and spinning on a latch is actually service time (and will be recorded as such). So it is entirely conceivable and demonstratable that a reduction in wait time will result in a larger increase in spin time. Although the wait time truly can be reduced, the more important response time can actually be increased!

Let's use Figure 5-12 as an example. Figure 5-12 is OraPub's interactive instance-level ORTA report. In summary, it gathers initial statistics, sleeps for the given interval (in this case 120 seconds), wakes and gathers statistics again, and reports their differences. Basic stuff, I know, but this report shows both service time and wait time, and sample active SQL during the report interval (not shown). Given the same workload (which is not shown on this earlier version of rtsysx.sql), our objective is to reduce the response time (shown as 305 seconds)-not just reduce the service time (shown as 119 seconds) or the wait time (shown as 186 seconds).

Figure 5-12. This OraPub 120-second interval response time report clearly shows Oracle processes are suffering from intense cache buffer chain latch contention. A response time-focused analysis will work on reducing the response time, not only the wait time or the service time.