I realize that most DBAs will not be so scientific in their work, but please do not forget to include some overhead when anticipating a parallelization increase. Forgetting or ignoring scalability will produce overly optimistic predictions.13

In Chapter 4, I noted that, in certain situations, it is desirable to encourage a system to be operating in the elbow of the response-time curve. That discussion should make a whole lot more sense now, and I'll summarize it again.

When focused on online user response time, a snappy response is desired. To increase the likelihood of snappy response time, we want the likelihood of queuing to be very low. We encourage this by ensuring a low utilization by keeping the arrival rate low enough to prohibit the system from operating in the elbow of the curve. While this produces snappy online response time, it also leaves available computing resources on the table. With a batch process focus, we want to use those leftover and available computing resources. In fact, to leave any available resource unused can be considered wasteful, shows parallelism is limited, and could result in a longer elapsed time. So, with batch processing, we look for ways to use any and all computing resources available to minimize elapsed time. This means the system will be operating in the elbow of the curve.