Compared to CPU, IO, and network analysis, memory bottleneck analysis is simple. The challenge is understanding the unique words used and not used on your particular platform. In fact, words like bottleneck and swapping can invoke a hostile and near-violent response from many operating system administrators. And don't even think about saying the word swap around an operating system vendor. It's just not worth the abuse you'll receive.

During one of my consulting engagements, it was clear that Oracle's memory requirements were exceeding the available capacity. In fact, the situation was so severe that it was impacting response times. When discussing the situation with the operating system administrator, I casually and gently said, "The system clearly has a memory bottleneck." He gave me a confused look and said, "We don't have a memory bottleneck. But there is definitely a lot of memory pressure." So I learned a new term that day to describe a memory bottleneck: memory pressure. It does make sense, and it is a kinder and gentler word, which I always try to use when situations are very tense.

Part of the confusion DBAs face stems from the fact that physical memory is used for a variety of purposes and classified differently. There is real memory, virtual memory, shared memory, nonshared memory, private memory, shared memory segments, code, data, stack, resident memory . . . I'm sure there are other types, names, and categories. The confusion can be significantly eliminated by grouping memory into three categories: real and virtual, shared memory segments, and process-related memory.