I both agree and disagree with your father. Performance should be thought about early, but micro-optimization should only be thought about early if you actually know that a high percent of time will be spent in small CPU-bound sections of code.

The problem with micro-optimization is that it is usually done without having any concept of how programs actually spend more time than necessary.

This knowledge comes from experience doing performance tuning, as in this example, in which a seemingly straightforward program, with no obvious inefficiencies, is taken through a series of diagnosis and speedup steps, until it is 43 times faster than at the beginning.

What it shows is that you cannot really guess or intuit where the problems will be. If you perform diagnosis, which in my case is random-pausing, lines of code responsible for a significant fraction of time are preferentially exposed. If you look at those, you may find substitute code, and thereby reduce overall time by roughly that fraction.

Other things you didn't fix still take as much time as they did before, but since the overall time has been reduced, those things now take a larger fraction, so if you do it all again, that fraction can also be eliminated. If you keep doing this over multiple iterations, that's how you can get massive speedups, without ever necessarily having done any micro-optimization.

After that kind of experience, when you approach new programming problems, you come to recognize the design approaches that initially lead to such inefficiencies. In my experience, it comes from over-design of data structure, non-normalized data structure, massive reliance on notifications, that sort of thing.

I both agree and disagree with your father. Performance should be thought about early, but micro-optimization should only be thought about early if you actually know that a high percent of time will be spent in small CPU-bound sections of code.

The problem with micro-optimization is that it is usually done without having any concept of how programs actually spend more time than necessary.

This knowledge comes from experience doing performance tuning, as in this example, in which a seemingly straightforward program, with no obvious inefficiencies, is taken through a series of diagnosis and speedup steps, until it is 43 times faster than at the beginning.

What it shows is that you cannot really guess or intuit where the problems will be. If you perform diagnosis, which in my case is random-pausing, lines of code responsible for a significant fraction of time are preferentially exposed. If you look at those, you may find substitute code, and thereby reduce overall time by roughly that fraction.

Other things you didn't fix still take as much time as they did before, but since the overall time has been reduced, those things now take a larger fraction, so if you do it all again, that fraction can also be eliminated. If you keep doing this over multiple iterations, that's how you can get massive speedups, without ever necessarily having done any micro-optimization.

After that kind of experience, when you approach new programming problems, you come to recognize the design approaches that initially lead to such inefficiencies. In my experience, it comes from over-design of data structure, non-normalized data structure, massive reliance on notifications, that sort of thing.