The situation begins with the system performing wonderfully. It's running so well that the DBA is asked to allow another batch job or two to run concurrently. What the DBA doesn't realize is that the system is operating just before the elbow of the curve. So the DBA decides to allow the increased workload to enter the system-and bam!-performance takes a significant drop. On a 12-CPU core system, it takes only a relatively small increase in the workload to send your system deep into the elbow of the curve. If the system were not operating near the elbow of the curve, then the situation would have been very different. This is perfect example of how understanding just a little queuing theory can make a big difference. If the DBA had known the number of cores, the CPU busyness, a little queuing theory, and some workload management control, he would have never allowed the increased workload. (I will delve deeper into practically applying queuing theory in the last chapter.)

Now let's shift to a CPU subsystem with just a couple of cores. Figures 4-5 and 4-6 are models based on a 2-CPU core system. If you graphically compare the 12-CPU core system with the 2-CPU core system, you will immediately notice a difference. The 12-CPU core system (Figure 4-4) response-time curve has a much more distinct elbow; that is, the system's response-time curve is very flat until it enters the elbow of the curve, and then response time skyrockets. In stark contrast, the 2-CPU core system graph (Figure 4-6) response-time curve elbow is more gradual. In fact, if you look at the queue time value in Figure 4-5, you'll notice that queue time exists even during a light workload. The situation is even more apparent with a single-CPU core system. This is why experienced DBAs know they can run a system with more cores at a higher CPU busyness than a system with just few cores!

Figure 4-5. Numerically, on a two-CPU core system, queuing accounts for 10% of response time at around 32% busy. Just as with Figure 4-3, when the system busyness exceeds 100%, queue time becomes infinite.