Figure 3-9. Latching acquisition performance data gathered from a real Oracle system, not based on a model. Notice that Oracle's default latching spin and sleep algorithm actually decreases latch acquisition response time per user call until the CPU subsystem becomes saturated. The test was run on a single four-CPU core Linux Oracle 10g Release 2 system.

Figure 3-9 provides deep insights into the efficiency of Oracle's latch acquisition algorithm. Initially, there is some overhead, resulting in a relatively high initial response time. This does not represent a problem, because there is plenty of CPU power available. As concurrency goes up and CPU cycles become increasingly less available, fortunately, economies of scale take effect, resulting in a continual decrease in acquisition service time and response time. But as CPU resources start to become scarce, at 47% CPU utilization, Oracle processes begin sleeping. However, even with the workload increasing, sleep time per user call remains about the same, and CPU time per user calls continues to decrease until around 95% utilization. At this point, CPU resources are in such short supply that queue time becomes increasingly more significant as processes spend more time sleeping.

This graph serves as a testimony to the efficiency of Oracle's latch acquisition algorithm, because even as the load increases pushing CPU utilization toward 100%, latch acquisition time does not significantly increase. If you look closely, only at 97% utilization does the average acquisition time begin to increase. Even more amazing is that, based on queuing theory, with four cores running at 80% utilization, around 50% of the response time is composed of queue time. Yet based on this test with Oracle's algorithm, queue time doesn't hit 50% of response time till 100% CPU utilization. So Oracle beat queuing theory!