How to Make Things Faster

Chapter 38. Every Problem Is a Skew Problem

I’m pretty sure that every performance problem I’ve ever helped solve can be framed as a skew problem. Even the label printer in Orange County. Skew is a nonuniformity in a list. It’s everywhere:

• Do all your task executions have uniform priority? No, some are more important than others. This is skew in program priorities. The principal motivation for creating Method R is that the skew in your diagnostic priorities needs to match the skew in your business priorities.

• Do all the programs on a system have the same profile? No, different programs usually have different profiles. For example, PYUGEN spent more time doing network I/O than anything else, but other programs on the same system spent more time using CPU and waiting for latches than anything else.

• Do all event types contribute uniformly to a profile? No, usually, some event types contribute more to total duration than others. You’ll probably never see a profile for a computer program with perfectly uniform contributions from all its event types.

• Do all the events of a given type contribute uniformly to a profile? No, some event executions take longer than other executions of the same event type. For example, your slowest hundred thousand disk read calls can contribute 100 times more duration than your fastest nine million calls.

• Does every execution of the same program have the same profile? No, a program might spend most of its time in the morning reading from SSD, and ...